Beruflich Dokumente
Kultur Dokumente
Applied Optimization
Methods for Wireless
Networks
Y. Thomas Hou
Yi Shi
Hanif D. Sherali
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9781107018808
Cambridge University Press 2013
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2013
Printed in the United Kingdom by TJ International Ltd. Padstow Cornwall
A catalog record for this publication is available from the British Library
Library of Congress Cataloguing in Publication Data
ISBN 978110701880
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication,
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
To our parents
and
Our wives Tingting, Meiyu, and Semeen
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Contents
Preface page xi
Acknowledgments xiv
Copyright permissions xvi
1 Introduction 1
1.1 Book overview 1
1.2 Book outline 3
1.3 How to use this book 7
viii Contents
ix Contents
x Contents
References 316
Index 327
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Preface
Reasons for writing the book In recent years, there is a growing trend
in applying optimization approaches to study wireless networks. Such an
approach is usually necessary when the underlying goal is to pursue funda-
mental performance limits or theoretical results. This book is written to serve
this need and is mainly targeted to graduate students who are conducting the-
oretical research in wireless networks using optimization-based approaches.
This book will also serve as a very useful reference for researchers who wish
to explore various optimization techniques as part of their research methodolo-
gies.
To prepare a graduate student in either electrical and computer engineer-
ing (ECE) or computer science (CS) to conduct fundamental research in wire-
less networks, an ideal roadmap would include a series of graduate courses in
operations research (OR) and CS, in addition to traditional communications
and networking courses in ECE. These OR and CS courses would include
(among others) linear programming, nonlinear programming, integer program-
ming from OR, and complexity theory and algorithm design and analysis
from CS. Today, these courses are typically offered as core courses within the
respective disciplines. Instructors in OR and CS departments typically have
little knowledge of wireless networks and are unable to make a connection
between the mathematical tools and techniques in these courses and problem-
solving skills in wireless networks. ECE/CS students often find it difficult to
see how these courses would benefit their research in wireless networks. Due
to this gap between teaching scopes and learning expectations, we find that
the learning experience of our ECE/CS students in these courses is passive (or
blind) at best, as they do not have a clear picture of how these courses will
benefit their research.
An approach to bridge this gap is to offer a course that reviews a collection of
mathematical tools from OR and CS (with a focus on optimization techniques)
and shows how they can be used to address some challenging problems in
wireless networks. This book is written for this purpose.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
xii Preface
Each chapter in this book starts with a brief pointer to the underlying opti-
mization technique (with references to tutorials or textbooks so that students
can do an in-depth study in a formal course or on their own). The chapter then
immediately delves into a detailed case study in wireless networks to which
the technique will be applied. The focus in each chapter is to show how the
underlying technique can be used to solve a challenging problem in wireless
networks. To achieve this goal, we offer details on how to formulate a research
problem into a formal optimization model, reformulate or transform it in order
to improve mathematical tractability, and apply the underlying optimization
technique (with necessary customizations that are specific to the underlying
problem) to derive an optimal or near-optimal solution.
We have taught this course a number of times to ECE and CS graduate stu-
dents at Virginia Tech, using chapters from this book. The response from the
students has been overwhelmingly positive. In particular, we find that:
xiii Preface
and Discrete Models (Athena Scientific, 1998). But that book still falls short
in showing students suitable case studies that are relevant to modern wireless
networks.
On the other hand, most other books addressing network optimization follow
a problem-oriented approach (vs. our method-oriented approach). They do not
offer a systematic treatment of the underlying optimization techniques like we
do in this book. To make this point clear, we quote the following text from the
preface of the book Combinatorial Optimization in Communication Networks,
edited by Maggie Xiaoyan Cheng, Yingshu Li, and Ding-Zhu Du (Springer,
2006), to explain why the problem-oriented approach was adopted by most
authors:
Two approaches were considered: optimization method oriented (starting from combi-
natorial optimization methods and finding appropriate network problems as examples)
and network problem oriented (focusing on specific network problems and seeking
appropriate combinatorial optimization methods to solve them). We finally decided to
use the problem-oriented approach, mainly because of the availability of papers: most
papers in the recent literature appear to address very specific network problems, and
combinatorial optimization comes as a convenient problem solver.
Acknowledgments
This book is the fruit of close collaboration among the three authors for more
than ten years. We would like first to thank the former and current members of
our research group. In particular, many thanks to: Jia (Kevin) Liu, whose work
led to Chapter 3, Sushant Sharma, whose work led to Chapter 4, Liguang Xie,
whose work led to Chapter 7, Dr. Shiwen Mao, whose work led to Chapter 11.
We want to thank Huacheng Zeng, Liguang Xie, Xu Yuan, and Canming Jiang
for their help in proofreading some of the chapters. They also contributed to the
preparation and revision of the solution manual and Powerpoint Slides. With-
out their help, this book would not have reached its current shape. Some other
former and current members of our group, whose names were not mentioned
above but who contributed to this book in many other ways, include Sastry
Kompella, Cunhao Gao, Tong Liu, Xiaojun Wang, Xiaolin Cheng, Dr. Rongbo
Zhu, Dr. Lili Zhang, and Dr. Wangdong Qi.
We also want to thank the students in our ECE/CS 6570 class (Advanced
Foundations of Networking) over the years, who offered valuable feedback
to different versions of this book and helped us gauge the best match of such
materials for a graduate course in networking. In particular, those students who
took ECE/CS 6570 in Fall 2012 directly contributed to proofreading the final
book manuscript and their feedback is greatly appreciated.
We want to thank Dr. Philip Meyler, Acquisitions Editor of Cambridge Uni-
versity Press, who showed a genuine interest in the initial conception of this
book and encouraged us to move forward for a book proposal. He has also
been extremely patient with us when we requested a one year extension for
the final delivery of our manuscript. We thank him for his trust, patience, and
understanding, which allowed us to work on our schedule to bring this book
to reality. Looking back, we feel really lucky that we chose the best publisher
for this book. During the manuscript preparation stage, we worked with three
different Assistant Editors of Cambridge University Press Elizabeth Horne,
Kirsten Bot, and Sarah Marsh. We thank all three of them, who worked dili-
gently with us at each step along the way to make this book a polished product.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
xv Acknowledgments
Tom Hou would like to thank Scott Midkiff, who recruited him to join the
Electrical and Computer Engineering Department at Virginia Tech in 2002.
Over the years, Scott has been a great colleague, a close friend, a resourceful
mentor, and, most recently, a supportive department head. The environment
that Scott and the department were able to offer has been instrumental to Toms
success in research and scholarship.
Finally, we would like to thank the National Science Foundation (NSF) and
the Office of Naval Research (ONR), whose funding support of our research
over the years led to this book.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Copyright permissions
Chapter 2
Y.T. Hou, Y. Shi, and H.D. Sherali, Rate allocation and network lifetime prob-
lems for wireless sensor networks, IEEE/ACM Transactions on Networking,
vol. 16, no. 2, pp. 321334, April 2008. Copyright 2008 by, and with kind
permission from, IEEE.
Chapter 3
J. Liu, Y.T. Hou, Y. Shi, and H.D. Sherali, Cross-layer optimization for
MIMO-based wireless ad hoc networks: routing, power allocation, and band-
width allocation, IEEE Journal on Selected Areas in Communications, vol. 26,
no. 6, pp. 913926, August 2008. Copyright 2008 by, and with kind permis-
sion from, IEEE.
Chapter 4
S. Sharma, Y. Shi, Y.T. Hou, and S. Kompella, An optimal algorithm for relay
node assignment in cooperative ad hoc networks, IEEE/ACM Transactions on
Networking, vol. 19, issue 3, pp. 879892, June 2011. Copyright 2011 by,
and with kind permission from, IEEE.
Chapter 5
Y. Shi, Y.T. Hou, and H. Zhou, Per-node based optimal power control
for multi-hop cognitive radio networks, IEEE Transactions on Wireless
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CHAPTER
1 Introduction
The present moment is the only moment available to us, and it is the door to
all moments.
Thich Nhat Hanh
1
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
2 Introduction
Combines techniques from both operations research (OR) and computer sci-
ence (CS) disciplines, with a strong focus on solving optimization problems
in wireless networks.
Shows various tricks and step-by-step details on how to develop optimiza-
tion models and reformulations, particularly in the context of cross-layer
optimization problems involving flow routing (network layer), scheduling
(link layer), and power control (physical layer).
Discusses case studies that focus on multi-hop wireless networks (e.g., ad
hoc and sensor networks) and incorporates a number of advanced physical
layer technologies such as MIMO, cognitive radio (CR), and cooperative
communications (CC).
Contains problem sets at the end of each chapter. PowerPoint slides for each
chapter are available to both the students and instructors. A solutions manual
is available to the instructors.
This book has four parts. Part I of this book, consisting of Chapters 2 to 4,
is devoted to optimization and designing algorithms that can offer optimal
solutions.
4 Introduction
6 Introduction
Part III, consisting of Chapters 10 and 11, is devoted to some highly effective
heuristics.
Part IV, currently consisting of only Chapter 12, is devoted to some miscel-
laneous topics in the broader context of wireless network optimizations. This
part will be further expanded to include other topics in a future book edition.
8 Introduction
PART
CHAPTER
The real voyage of discovery consists not in seeking new landscapes but in
having new eyes.
Marcel Proust
11
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
The fairness issue associated with the network capacity maximization objec-
tive calls for careful consideration in rate allocation among the nodes. In this
chapter, we study the rate allocation problem in an energy-constrained sen-
sor network for a given network lifetime requirement. Our objective is to
achieve a certain measure of optimality in the rate allocation that takes into
account both fairness and bit rate maximization. We employ the so-called Lex-
icographic Max-Min (LMM) criterion [99], which maximizes the bit rates
for all the nodes under the given energy constraint and network lifetime
requirement. At the first level, the smallest rate among all the nodes is max-
imized. Then, we continue to maximize the second smallest rate level and so
forth. The LMM rate allocation criterion is appealing since it addresses both
fairness and efficiency (i.e., bit rate maximization) in an energy-constrained
network.
A naive approach to the LMM rate allocation problem would be to apply
a max-min-like iterative procedure. Under this approach, successive LPs are
employed to calculate the maximum rate at each level based on the avail-
able energy for the remaining nodes, until all nodes use up their energy. We
call this approach serial LP with energy reservation (SLP-ER). We will
show that, although SLP appears intuitive, it is likely to offer an incorrect
solution. To understand how this could happen, we must understand a funda-
mental difference between the LMM rate allocation problem described here
and the classical max-min rate allocation in [14]. Under the LMM rate allo-
cation problem, the rate allocation is implicitly coupled with a flow routing
problem, while under the classical max-min rate allocation, there is no rout-
ing problem involved since the routes for all flows are given. As it turns out,
for the LMM rate allocation problem, any iterative rate allocation approach
that requires energy reservation at each iteration is incorrect. This is because,
unlike max-min, which addresses only the rate allocation problem with fixed
routes and yields a unique solution at each iteration, for the LMM rate allo-
cation problem, there usually exist nonunique flow routing solutions corre-
sponding to the same rate allocation at each level. Each of these flow routing
solutions will yield different available energy levels on the remaining nodes
for future iterations and so forth. This will lead to a different rate alloca-
tion vector, which may not coincide with the optimal LMM rate allocation
vector.
In this chapter, we show a correct approach to solve the LMM rate alloca-
tion problem. Our approach employs the so-called parametric analysis (PA)
technique [10] in LP to determine the minimum set of nodes at each rate level
that must deplete their energy. We call this approach serial LP with PA (SLP-
PA). We also extend the PA technique for the LMM rate allocation problem to
address the so-called maximum node lifetime curve problem in [20], which
we call the LMM node lifetime problem. More importantly, we show that
there exists a simple and elegant mirror relationship between the LMM rate
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
allocation problem and the LMM node lifetime problem. As a result, it is suf-
ficient to solve only one of these two problems.
The remainder of this chapter is organized as follows. Section 2.3 intro-
duces the network and energy model, and formulates the LMM rate alloca-
tion problem. Section 2.4 describes the SLP-PA algorithm for the LMM rate
allocation problem. Section 2.5 introduces the LMM node lifetime problem
and shows how to apply the SLP-PA algorithm to solve it. Section 2.6 reveals
an interesting mirror relationship between the LMM rate allocation problem
and the LMM node lifetime problem. Section 2.7 presents numerical results.
Section 2.8 summarizes this chapter.
Base station
(BS)
Upper tier
AFN
Lower tier
MSN
General notation for the LMM rate and LMM lifetime problems
where 1 and 2 are two constant terms, dij is the distance between these
two nodes, and is the path-loss index. Typical values for these param-
eters are 1 = 50 nJ/b, 2 = 0.0013 pJ/b/m4 [67], and = 4 [129]. Since
the power level of an AFNs transmitter can be used to control the distance
coverage of an AFN (e.g., [127; 131; 166]), different network flow rout-
ing topologies can be formed by adjusting the power level of each AFNs
transmitter.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
The set of constraints in (2).5 are the flow balance equations. They state that
the total bit rate transmitted by AFN i is equal to the total bit rate received by
AFN i from other AFNs, plus the bit rate generated locally at this node (ri ).
Note that we allow flow splitting at a node so as to achieve more flexibility
in flow routing and load balancing in the network. The set of constraints in
(2).6 are the energy constraints. They state that for a given network lifetime
requirement T , the energy consumed for communications (i.e., transmitting
and receiving) cannot exceed the initial available energy.
Note that fki , fij , fiB , and ri are variables and T is a constant (the
given network lifetime requirement). Since MaxCap is an LP problem, it
can be solved in polynomial time [10]. Unfortunately, as we will see in the
numerical results (Section 2.7), the solution to this MaxCap problem favors
those AFNs whose data paths consume the least amount of power toward
the base station. Consequently, although the sum of rates is maximized over
T , the specific bit rate allocation among the AFNs (i.e., ri s) favors those
AFNs that have this property, while the other AFNs are unfavorably allocated
with much smaller (even close to 0) data rates. As a result of this unfair-
ness, the effectiveness of the sensor network in performing data collection is
questionable.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Definition 2.1
LMM-optimal rate allocation For a given network lifetime require-
ment T , a sorted rate vector g = [g1 , g2 , . . . , gN ] yields an LMM-optimal
rate allocation if and only if for any other sorted rate allocation vector
g = [g1 , g2 , . . . , gN ] with g1 g2 . . . gN there exists k, 1 k N ,
such that gi = gi for 1 i k 1 and gk > gk .
Maximize 1
subject to fiB + j =i fij k=i fki 1 = 0 (1 i N )
k=i Tfki + j =i Cij Tfij + CiB TfiB ei (1 i N )
fij , fiB 0 (1 i, j N, j = i).
Once we obtains a solution with the maximum 1 , we can also calculate the
energy consumption at each node under this flow routing solution. Then we can
check whether or not a node has any remaining energy. If there are some nodes
that still have remaining energy, then we can construct another LP problem
to further increase their data rates (with a maximum rate increment of 2 ).
1 However, there is a significant difference between max-min and LMM, which we will discuss
shortly.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
This process terminates until all nodes use up their energy. Since energy is
reserved at each node after each iteration, we call this naive approach Serial
LP with Energy Reservation (SLP-ER).
SLP-ER cannot provide an LMM-optimal solution because there is a funda-
mental difference in the nature of the LMM rate allocation problem described
here and the classical max-min rate allocation problem in [14]. The LMM rate
allocation problem implicitly couples a flow routing problem (i.e., a determi-
nation of the fij and fiB for the entire network), while the classical max-min
rate allocation explicitly assumes that the routes for all the flows are given a
priori and remain fixed. For the LMM rate allocation problem, starting from
the first iteration, there usually exist nonunique flow routing solutions corre-
sponding to the same maximum rate level. Consequently, each of these flow
routing solutions, once chosen, will yield different remaining energy levels on
the nodes for future iterations and so forth. This will lead to a nonunique rate
allocation vector, which may not coincide with the LMM-optimal rate vector
(see numerical examples in Section 2.7).
as the corresponding set of nodes that use up their energy at this rate. Clearly,
we have ni=1 |Si | = |S| = N , where S denotes the set of all N nodes.
The essence to solving the LMM rate allocation problem is to find the correct
values for 1 , 2 , . . . , n and the corresponding sets S1 , S2 , . . . , Sn , respec-
tively. This can be done iteratively. We first determine the rate level 1 and the
corresponding set S1 , then determine the rate level 2 and the corresponding
set S2 , and so on. In Section 2.4.1, we will show how to determine each rate
level and in Section 2.4.2, we will show how to determine the corresponding
node set.
Maximize l
l1
subject to fiB + j =i fij k=i fki l = l1 (i Sh ) (2.7)
h=0
fiB + j =i fij k=i fki = h (i Sh , 1 h < l) (2.8)
l1
( k=i fki + j =i Cij fij + CiB fiB )T ei (i Sh ) (2.9)
h=0
( k=i fki + j =i Cij fij + CiB fiB )T = ei (i Sh , 1 h < l) (2.10)
fij , fiB 0 (1 i, j N, j = i).
Note that for l = 1, constraints (2).8 and (2).10 do not exist. For 2 l n,
constraints (2).8 and (2).10 are for those nodes that have already reached their
LMM rate allocation during the previous (l 1) iterations. In particular, the
set of constraints in (2).8 say that the sum of in-coming and local data rates
is equal to the out-going data rates for each node with its LMM-optimal rate
h , 1 h < l. The set of constraints in (2).10 say that for those nodes that
have already reached their LMM-optimal rates, the total energy consumed for
communications has reached their initial energy provisioning. On the other
hand, the constraints in (2).7 and (2).9 are for the remaining nodes that have
not yet reached their LMM-optimal rates. Specifically, the set of constraints in
(2).7 state that, for those nodes that have not yet reached their energy constraint
levels, the sum of in-coming and local data rates is equal to the out-going data
rates. Note that the objective function is to maximize the additional rate l for
those nodes. Furthermore, for those nodes, the set of constrains in (2).9 state
that the total energy consumed for communications should be upper bounded
by the initial energy provisioning.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
LMM-rate Maximize l
l1
subject to ViB + j =i Vij k=i Vki l T = l1 T (i Sh ) (2.11)
h=0
ViB + j =i Vij k=i Vki = h T (i Sh , 1 h < l)
l1
Although a solution to the LMM rate problem gives the optimal solution for
l at iteration l, it remains to determine the minimum set of nodes correspond-
ing to this l , which is the key difficulty in the LMM rate allocation problem.
Fortunately, this problem can be solved by the so-called parametric analysis
(PA) technique in LP [10].
Lemma 2.1
The minimum node set for each rate level under the LMM-optimal rate
allocation is unique.
With this lemma in place, we now discuss how to determine the minimum
node set Sl corresponding to the rate level l . Denote Sl ( = ) as the set of
nodes for which the constraints (2).9 are binding at the lth iteration in LMM
rate, i.e., Sl include all the nodes that achieve equality in (2).9 at iteration l.
Although it is certain that at least one of the nodes in Sl belong to Sl (the
minimum node set for rate l ), some nodes in Sl may still achieve greater rates
under other flow routing solutions. In other words, if |Sl | = 1, then we must
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Lemma 2.2
The optimal flow routing solution corresponding to the LMM rate alloca-
tion may not be unique.
Example 2.1
Consider an eight-node network with a topology shown in Fig. 2.2. The
base station B is located at the origin (0, 0). There are two groups of nodes,
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
G1 and G2 , in the network, with each group consisting of four nodes. Group
G1 consists of AFN1 at (100, 0), AFN3 at (0, 100), AFN5 at (100, 0),
and AFN7 at (0, 100), respectively (all in meters); Group G2 consists of
AFN2 at (100, 100), AFN4 at (100, 100), AFN6 at (100, 100), and
AFN8 at (100, 100), respectively. Assume that all nodes have the same
initial energy e. For a network lifetime requirement of T , we can calculate
(via SLP-PA) that the final LMM-optimal rate allocation for all eight nodes
are identical (perfect fairness), i.e., r1 = r2 = . . . = r8 . We denote ri = g
for 1 i 8.
Y (m)
4 39g/62 x 3 x 2
100
23
39g/62 x
101g/ 62
g/
62
62
g/
x
23
5 101g/ 62 101g/ 62 1
X (m)
100 0 100
101g/ 62
39g/62 x
23
62
g/ x
g/
62
23
x 39g/62 x
6 100
7 8
Figure 2.2 A simple example showing that the optimal flow routing to the LMM rate allocation is not
unique. The range of x is 0 x 39g
62 .
In this section, we show that the SLP-PA algorithm can be used to solve the
so-called maximum node lifetime curve problem in [20], which we define as
the LMM node lifetime problem. We also show that the SLP-PA algorithm is
much more efficient than the one proposed in [20].
Definition 2.2
LMM-optimal node lifetime A sorted node lifetime vector
[1 , 2 , . . . , N ] with 1 2 . . . N is LMM-optimal if and only if for
any other sorted node lifetime vector [1 , 2 , . . . , N ], with 1 2 . . .
N , there exists a k, 1 k N such that i = i for 1 i k 1 and
k > k .
2.5.2 Solution
It should be clear that, under the LMM-optimal node lifetime objective, we
must maximize the time until a set of nodes use up their energy (which is also
called a drop point in [20]), while minimizing the number of nodes that drain
up their energy at each drop point. We now show that the SLP-PA algorithm
developed for the LMM rate allocation problem can be applied to solve the
LMM node lifetime problem.
Suppose that [1 , 2 , . . . , N ] with 1 2 . . . N is LMM-optimal.
To keep track of distinct node lifetimes (or drop points) in this vector, we
remove all repetitive elements in the vector and rewrite it as [1 , 2 , . . . , n ]
such that 1 < 2 < . . . < n , where 1 = 1 , n = N , and n N . Corre-
sponding to these drop points, denote S1 , S2 , . . . , Sn as the sets of nodes that
drain up their energy at drop points 1 , 2 , . . . , n , respectively. Then |S1 | +
|S2 | + . . . + |Sn | = |S| = N, where S denotes the set of all N AFNs in the net-
work. The problem is to find the LMM-optimal values of 1 , 2 , . . . , n and
the corresponding sets S1 , S2 , . . . , Sn .
Similar to the LMM rate allocation problem, the LMM node lifetime prob-
lem can be formulated as an iterative optimization problem as follows. Denote
0 = 0, S0 = , and l = l l1 . Starting from l = 1, we solve the fol-
lowing LP problem iteratively:
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Comparing the above LMM lifetime problem to the LMM rate problem in
Section 2.4.1, we find that they are mathematically identical. The only difference
is that under the LMM lifetime problem, the local bit rates ri s are constants and
the node lifetimes i s are variables, while under the LMM rate problem, the
ri s are variables and the node lifetimes are all identical (T ), i = 1, 2, . . . , N.
Since the mathematical formulation for the two problems are identical, the
SLP-PA algorithm can be applied to solve the LMM node lifetime problem.
The only issue that we need to be concerned about is the optimal flow rout-
ing solution corresponding to the LMM-optimal lifetime vector. The optimal
flow routing solution here is not as simple as that for the LMM rate allocation
problem, which merely involves a simple division (see (2).12). We refer read-
ers to the appendix at the end of this chapter for an O(N 4 ) algorithm to obtain
an optimal flow routing solution for the LMM-optimal lifetime vector. Similar
to Lemma 2.2, the optimal flow routing solution corresponding to the LMM
node lifetime problem may not be unique.
In this section, we present an elegant result showing that there is a mirror rela-
tionship between the LMM rate allocation problem and the LMM node lifetime
problem. As a result, it is only necessary to solve only one of the two problems
and the results for the other problem can be obtained via simple algebraic cal-
culations.
To start with, we denote PR as the LMM rate allocation problem where we
have N AFNs in the network and all nodes have a common lifetime require-
ment T (a given constant). Denote ri as the LMM-optimal rate allocation for
node i under PR , i = 1, 2, . . . , N. Similarly, we denote PL as the LMM node
lifetime problem where all nodes have the same local bit rate R (constant).
Denote ti as the LMM node lifetime for node i under PL , i = 1, 2, . . . , N.
Then the following theorem shows how the solution to one problem can be
used to obtain the solution to the other.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 2.2 Mirror relationship between the LMM rate allocation problem PR and the
LMM node lifetime problem PL .
PR PL
Theorem 2.1
Mirror relation For a given node lifetime requirement T for all nodes
under Problem PR and a given local bit rate R for all nodes under Problem
PL , we have the following relationship between the solutions to the LMM
rate allocation problem PR and the LMM node lifetime problem PL .
(i) Suppose that we have solved Problem PR and obtained the LMM-
optimal rate allocation ri for each node i (i = 1, 2, . . . , N ). Then under
PL , the LMM node lifetime ti for node i is given by
ri T
ti =. (2.14)
R
(ii) Suppose that we have solved Problem PL and obtained the LMM-
optimal node lifetime ti for each node i (i = 1, 2, . . . , N). Then under PR ,
the LMM rate allocation ri for node i is given by
ti R
ri = . (2.15)
T
Now, replacing ri T by ti R, we see that the same bit volume solution under PR
yields a feasible bit volume solution to the node lifetime problem under PL .
Consequently, we can use Algorithm 2.1 to obtain the flow routing solution to
Problem PL under the bit volume solution to Problem PL and this verifies that
ti , i = 1, 2, . . . , N, is a feasible solution to Problem PL .
Optimality The proof is based on contradiction. But first, lets give some
notations. To prove that the ti -values obtained via (2).14 are indeed LMM-
optimal for Problem PL , we sort ri , i = 1, 2, . . . , N , under Problem PR in
nondecreasing order and denote it as [g1 , g2 , . . . , gN ]. We also introduce
a node index I = [i1 , i2 , . . . , iN ] for [g1 , g2 , . . . , gN ]. For example, i3 = 7
means that g3 actually corresponds to the rate of AFN 7, i.e., g3 = r7 .
Since ti is proportional to ri through the relationship (ti = R T
ri ), listing
ti , i = 1, 2, . . . , N, according to I = [i1 , i2 , . . . , iN ] will yield a sorted (in
nondecreasing order) lifetime list, denoted as [1 , 2 , . . . , N ]. We now prove
that if [1 , 2 , . . . , N ] is not LMM-optimal for Problem PL , then we will have
that [g1 , g2 , . . . , gN ] is not LMM-optimal, which is a contradiction.
Suppose that [1 , 2 , . . . , N ] is not LMM-optimal for Problem PL . Assume
that the LMM-optimal lifetime vector to Problem PL is [1 , 2 , . . . , N ]
(sorted in non-decreasing order) with the corresponding node index being
I = [i1 , i2 , . . . , iN ]. Then, by Definition 2.2, there exists a k such that j = j
for 1 j k 1 and k > k .
We now claim that if ti , i = 1, 2, . . . , N, is a feasible solution to Problem
PL , then ri obtained via ri = tiTR , i = 1, 2, . . . , N, is also a feasible solution
to Problem PR . The proof to this claim follows identically as above. Using
this result, we can obtain a corresponding feasible solution [g1 , g2 , . . . , gN ]
with gi = Ti R and the node index I for Problem PR . Hence, we have
R R
gj = jT = jT = gj for 1 j k 1 but gk = kTR > kTR = gk . That is,
[g1 , g2 , . . . , gN ] is not LMM-optimal and this leads to a contradiction.
(ii) The proof for this part is similar to that for (i) and is left as a homework
exercise.
Y (m) Y (m)
500 500
X (m) X (m)
500 0 500 500 0 500
500 500
Table 2.5 Rate allocation under three approaches for the 10-AFN network.
3.5
1.2
3 SLPPA SLPPA
SLPER 1 SLPER
2.5 MaxCap MaxCap
Rate (kb/s)
0.8
Rate (kb/s) 2
0.6
1.5
1 0.4
0.5 0.2
0 0
2 4 6 8 10 12 14 16 18 20 1 2 3 4 5 6 7 8 9 10
Sorted node index Sorted node index
(a) A 10-AFN network. (b) A 20-AFN network.
Table 2.6 Rate allocation under three approaches for the 20-AFN network.
(see Section 2.6). Again, we use the 10-AFN and 20-AFN network configura-
tions in Figs. 2.3(a) and (b), respectively. For both networks, we assume that
the initial energy at each AFN is 50 KJ and that the network lifetime require-
ment under the LMM rate allocation problem is T = 100 days. Under PL , we
assume that the local bit rate for all AFNs is R = 0.2 Kb/s.
To verify the mirror relationship (Theorem 2.1), we perform the following
calculations. First, we solve the LMM rate allocation problem (PR ) and the
LMM node lifetime problem (PL ) independently with the above initial condi-
tions using the SLP-PA algorithm. Consequently, we obtain the LMM-optimal
rate allocation (ri for each AFN i) under PR and the LMM-optimal node life-
time (ti for each AFN i) under PL . Then we compute T ri and R ti sepa-
rately for each AFN i and examine if they are equal to each other.
The results for the LMM-optimal rate allocation (ri , i = 1, 2, . . . , 10) and
the LMM-optimal node lifetime (ti , i = 1, 2, . . . , 10) for the 10-AFN network
are shown in Table 2.7. We find that T ri and R ti are exactly equal for all
AFNs, precisely as we would expect under Theorem 2.1. Similarly, the results
for the 20-AFN network are shown in Table 2.8.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 2.7 Mirror relationship T ri D R ti between the LMM rate allocation prob-
lem (PR ) and the LMM node lifetime problem (PL ) for the 10-AFN network.
Table 2.8 Mirror relationship T ri D R ti between the LMM rate allocation prob-
lem (PR ) and the LMM node lifetime problem (PL ) for the 20-AFN network.
where the fki(l) -values, if not zero, have all been defined before calculating
the flow routing for node i.
4. Let Ul = Ul {i} and go to Step 2.
(5) (5)
out-going flow at node 2 is routed as follows: f24 = 0.4 Kb/s, f25 = 0.8
(5)
Kb/s, and f2B = 1.2 Kb/s.
2.9 Problems
37 2.9 Problems
numerical result to illustrate this point. You may use numbers with normalized
units (dimensionless) in this numerical example.
2.4 Suppose that we obtain the following three sorted rate vectors corre-
sponding to three feasible rate allocations (each following a different solution
approach):
a = [0.10, 0.10, 0.10, 0.15, 0.29, 0.29, 0.29, 0.29, 0.29, 0.29]
b = [0.10, 0.10, 0.10, 0.10, 0.10, 0.15, 0.15, 0.15, 0.65, 0.65]
c = [0.05, 0.06, 0.06, 0.06, 0.12, 0.16, 0.17, 0.26, 0.35, 1.23]
Which two vectors are definitely not LMM-optimal? Why?
2.5 Discuss the similarity and the difference between the max-min rate allo-
cation problem and the LMM rate allocation problem.
2.6 In this chapter, two incorrect solution approaches for the LMM rate allo-
cation are presented. Explain why these two approaches cannot provide an
LMM-optimal rate allocation.
2.7 What is the objective of using parametric analysis (PA) in the solution of
the LMM rate allocation problem? Explain the basic idea of PA.
2.8 Is the minimum node set for each rate level under the LMM-optimal rate
allocation unique? Is the flow routing solution corresponding to the LMM-
optimal rate allocation unique?
2.9 What is the connection between the LMM rate allocation problem and
the LMM node lifetime problem in terms of mathematical formulation? What
is the mirror relationship between the two problems and when does it occur?
What is the application of this mirror relationship?
2.10 For Theorem 2.1, give a proof for the second statement:
Suppose that we have solved Problem PL and obtained the LMM-optimal
node lifetime ti for each node i (i = 1, 2, . . . , N). Then under PR , the LMM
rate allocation ri for node i is
ti R
ri = .
T
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CHAPTER
Virgil
1. Convex set: A set is convex if for any two of its elements z1 and z2 and for
any [0, 1], z1 + (1 )z2 is also an element of this set. For example,
the set {(x, y) : x 2 + y 2 1} is a convex set but the set {(x, y) : 1 x 2 +
y 2 2} is not a convex set.
2. Convex and concave functions: A function f (x) is a convex function
if for any x1 and x2 and any [0, 1], f (x1 + (1 )x2 ) f (x1 ) +
(1 )f (x2 ), where x can be a single variable or a vector of variables. A
function f (x) is a concave function if for any x1 and x2 and any [0, 1],
f (x1 + (1 )x2 ) f (x1 ) + (1 )f (x2 ). For example, f (x) = x 2 is
a convex function, f (x) = ln x is a concave function, and f (x) = x 3 is
neither convex nor concave. Observe that if f (x) is a convex function, then
f (x) is a concave function, and vice versa.
3. Affine function: A function f (x) is an affine function if it is both convex
and concave. It can be verified that an affine function is a linear function
plus some real constant, i.e., f (x) = ax + b for some constants a and b.
38
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CP Maximize f (x)
subject to g(x) 0
h(x) = 0
x X,
LD Minimize (u, v)
subject to u 0,
where u is the Lagrangian multiplier (or dual variable) associated with the
constraint g(x) 0, the unconstrained variable v is the Lagrangian multiplier
associated with the constraint h(x) = 0, and where
The function f (x) uT g(x) vT h(x) is called the Lagrangian function and
provides an upper bound on f (x) for any feasible solution to CP. Hence,
max {f (x)}
xX,g(x)0,h(x)=0
for any u 0, and therefore v(LD) v(CP ), where v(p) denotes the optimal
value for any problem p. This latter relationship is known as the weak dual-
ity property. Under certain constraint qualifications (e.g., Slaters condition,
which requires the existence of an x X such that h(x) = 0, g(x) < 0, and the
origin belongs to the interior of the set {h(x) : x X}, convex programming
problems have the strong duality relationship v(LD) = v(CP ), i.e., both the
original primal problem and its dual problem have the same optimal objective
value [11]. Moreover, given an optimal solution to one problem, the optimal
solution to the other can be constructed using a suitable algorithmic proce-
dure [11]. Therefore, a primal problem can be possibly solved in its dual
domain.
An excellent reference on convex optimization is [11]. In the rest of this
chapter, we offer a case study to show how convex optimization can be applied
to solve a problem for a multi-hop MIMO network.
where O (n) is the set of outgoing links that from node n and
Pmax is the max-
Symbol Definition
c(Wl , Ql ) Wl log2 det I + l Hl Ql Hl , (3.1)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Similarly, we assume the sum of the bandwidth of all incoming links at node n
cannot exceed BI(n) , the bandwidth at this node that is allocated for reception.
We have
Wl BI(n) (1 n N ).
lI(n)
T
We denote the vector W = W1 W2 . . . WL RL as the collection of all
bandwidth variables.
It is easy to see that the values of Wl and Ql , i.e., the allocation of bandwidth
and power on link l, directly affect the capacity of link l. The determination of
these values will be part of our cross-layer optimization problem.
We assume there are F sessions in the network. We denote the source and
destination nodes of session f , 1 f F , as s(f ) and d(f ), respectively.
We allow flow splitting inside the network so as to offer more flexibility in
flow routing and better load balancing. Denote rf as the flow rate of session
f , 1 f F , which are optimization variables.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
(f )
Denote yl 0 as the amount of flow on link l attributed to session f . For
each session f , we have the following flow balance at node n:
(f ) (f ) rf if n = s(f ),
yl yl = rf if n = d(f ), (3.2)
lO(n) lI(n) 0 otherwise,
where I (n) is the set of links that are incoming to node n.
To rewrite (3.2) compactly, we use Rf RN to denote the right-hand-side
(RHS) of constraints for a session f , i.e., (Rf )s(f ) = rf , (Rf )d(f ) =
rf , and (Rf )n = 0 for 1 n N, n = s(f ), n = d(f ). We define y(f ) =
[ y1(f ) y2(f ) . . . yL(f ) ]T RL as the flow vector on all L links attributed to ses-
sion f . With NLIM, for any session f , (3.2) can be written Ay(f ) = Rf .
as NF
In addition, denoting the matrix R R1 R2 . . . RF R as the col-
lection of all source-sink vectors Rf , we have Rf = Ref . The matrix R needs
to satisfy the following constraints:
(Ref )s(f ) = rf (1 f F )
1, Ref = 0 (1 f F )
Ref = s(f ),d(f ) 0 (1 f F ),
where the notation =x,y represents the component-wise equality of a vector
except at the xth and yth entries. Denote the matrix Y y(1) y(2) . . . y(F )
RLF as the collection of all flow vectors y(f ) . Then, the flow conservation law
for all sessions can be written as
AY = R.
Since all flows traversing a link cannot exceed this links capacity limit,
(f )
we have Ff =1 yl c(Wl , Ql ) for 1 l L. Using matrix-vector notation,
this can be compactly written as
1, YT el c(Wl , Ql ) (1 l L).
Note that link capacity c(Wl , Ql ) depends on bandwidth allocation and the
input covariance matrix. Also note that this constraint couples three layers (i.e.,
network, link, and physical).
Wl BO(n) (1 n N ) (3.4)
lO(n)
Wl BI(n) (1 n N ) (3.5)
lI(n)
Minimize (u)
subject to u 0,
where
F
L
(u) = max ln rf + ul c(Wl , Ql ) 1, YT el .
(R,Y,Q,W)X
f =1 l=1
(3.11)
It is easy to recognize that, for any given Lagrangian multiplier u, the
Lagrangian can be separated into two terms:
linkphy (u) Maximize l ul c(Wl , Ql )
subject to lO(n) Tr{Ql } Pmax (1 n N )
W BO(n) (1 n N )
lO(n) l
W
lI(n) l B I(n) (1 n N )
Ql
0, Wl 0 (1 l L).
For the transport-network layer subproblem, we can regard f ln rf as the
revenue (measured by a rate utility function) and l ul 1, YT el as the cost to
achieve session rates. For the link-physical layer subproblem, we can regard
l ul c(Wl , Ql ) as the revenue, which is measured by a sum of weighted link
capacities. Then, the Lagrangian dual problem of P can be transformed into
the following problem:
D: Minimize tsptnet (u) + linkphy (u)
subject to u 0.
The Lagrangian dual problem requires us to find the optimal u such that
tsptnet (u) and linkphy (u) can be minimized. Thus, we should be able
to evaluate the subproblem functions tsptnet (u) and linkphy (u) for any
given u. Note that to evaluate tsptnet (u) in the network layer subproblem,
the objective function is concave and all constraints are affine. Hence, this
subproblem is a convex program, as described in Section 3.1. Following the
same token, we can verify that the problem of evaluating linkphy (u) is also a
convex program. Therefore, both tsptnet (u) and linkphy (u) can be readily
solved by many efficient convex programming methods. In the next section,
we show how to solve Problem D.
the relaxed problem, whose solution provides a lower bound for Problem D.
If this solution is not feasible to Problem D, we generate an additional linear
constraint to cut the current solution. With this additional linear constraint,
we obtain a tighter linear relaxation for the next iteration. Eventually, if we
obtain a linear relaxation using a finite number of constraints whose solution
to Problem D is feasible, then we would have derived an optimal solution to
Problem D.
We now present the details of the cutting-plane method. Let z = (u) =
F L
f =1 ln rf + l=1 ul c(Wl , Ql ) 1, Y el . Thus,
max(R,Y,Q,W)X T
for each u 0, the linear inequality z Ff =1 ln rf + L l=1 ul
c(Wl , Ql ) 1, Y el must hold for all (R, Y, Q, W) X. Problem
T
D is therefore equivalent to
Minimize z
subject to z f ln rf + l ul c(Wl , Ql ) 1, YT el , (3.12)
(R, Y, Q, W) Xu 0.
Note that each given point (R, Y, Q, W) yields a linear constraint for z and u.
Therefore, we can regard (3.12) as an LP problem having an infinite number of
constraints, with each constraint corresponding to a point (R, Y, Q, W) X.
Instead of enumerating all (infinite number of) points, we consider only a
finite number of points and add more points as needed. This gives us a relaxed
problem (since we consider only a subset of constraints). In the first iteration,
we begin with one point (R(0) , Y(0) , Q(0) , W(0) ) X.
In the kth iteration, we have k points (R(j ) , Y(j ) , Q(j ) , W(j ) ) X, j =
0, . . . , k 1, for which the relaxed problem is given as follows:
Minimize z
(j ) (j ) (j )
subject to z f ln rf + l ul c(Wl , Ql ) 1, Y(j )T el
(0 j < k) u 0.
(3.13)
Note that this problem is exactly the problem in (3.11). Therefore, as discussed
in Section 3.4, we can decompose this problem into two smaller subproblems
and solve each subproblem efficiently.
Denote (R(k) , Y(k) , Q(k) , W(k) ) as an optimal solution to (3.14) and (u(k) )
as the corresponding optimal objective value. If zk (u(k) ), then (z(k) , u(k) )
is feasible for (3.12) and is an optimal solution to this Lagrangian dual prob-
lem. Otherwise, for (z(k) , u(k) ), the inequality constraint in (3.12) is not satis-
fied for (R(k) , Y(k) , Q(k) , W(k) ). Thus, we add the constraint:
(k) (k) (k)
z ln rf + ul c(Wl , Ql ) 1, Y(k)T el (3.15)
f l
to (3.13), and solve the relaxed linear program again in the (k + 1)th itera-
tion (now with (k + 1) points). Obviously, (z(k) , u(k) ) will not show up in the
(k + 1)th iteration because it is cut off by (3.15). The cutting-plane algorithm
is summarized in Algorithm 3.1.
Thus far, we have solved Problem D. We now show how we can construct an
optimal solution to Problem P.
Suppose that Algorithm 3.1 terminates at iteration k = K. Then we have
K + 1 points (R(j ) , Y(j ) , Q(j ) , W(j ) ) X, j = 0, . . . , K. The last solved LP
problem in Algorithm 3.1 is as follows:
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Minimize z
(j ) (j ) (j )
subject to z f ln rf + l ul c(Wl , Ql ) 1, Y(j )T el
(0 j K) u 0.
(3.16)
As discussed in Section 2.1, when we solve (3.16), we have also solved its dual
problem, which is given as follows:
F (j )
Maximize K j =0 j f =1 ln(rf )
K
(j ) (j )
subject to j =0 j 1, Y(j )T el c(Wl , Ql ) 0, l
K (3.17)
j =0 j = 1
j 0, j.
By the LP duality theory, the optimal values j , for j = 0, . . . , K, are imme-
diately available after we solve (3.16). Furthermore, (3.16) and (3.17) have the
same optimal objective value z(K) .
To construct an optimal solution to Problem P, we first need to construct a
feasible solution. Such a feasible solution (R, Y, Q, W) should satisfy (3.10)
and (R, Y, Q, W) X. Since j , for j = 0, . . . , K, is a solution of (3.17), we
have K j =0 j (R , Y , Q , W ) X. Let
(j ) (j ) (j ) (j )
K
(R, Y, Q, W) = j (R(j ) , Y(j ) , Q(j ) , W(j ) ).
j =0
Theorem 3.1
(R, Y, Q, W) is a feasible and optimal solution to Problem P.
800 N7 N5 N6
N11 N9
N10
(m)
200 N1
0
0 200 400 600 800 1000 1200
(m)
30 MHz, respectively. There are three sessions in the network: node 14 to node
1, node 6 to node 10, and node 5 to node 4. The randomly generated channel
gain matrices are shown in Table 3.2.
After executing our solution procedure, we find that the optimal rates for
these three sessions are r1 = 125 Mb/s, r2 = 190.6 Mb/s, and r3 = 258 Mb/s.
The optimal objective value is 6.64. The routings and flow rates of sessions 1,
2, and 3 are shown in Figs. 3.2, 3.3, and 3.4, respectively. These figures show
that flow routings for sessions 1, 2, and 3 are all multi-path and multi-hop. It
can be easily verified that the flow rates in Figs. 3.2, 3.3, and 3.4 satisfy flow
conservation at each node.
Table 3.3 shows the optimal bandwidth allocation of the network. Table 3.4
shows the optimal input covariance matrices of the network. In Table 3.4, each
cell with four entries corresponds to a 2 2 Q-matrix, which represents an
input covariance matrix. Consider the transmission from N11 to N7 as an
example: Q(11,7) in Table 3.4 says that the power allocations to the two anten-
nas at N11 are 11.83 mW and 12.10 mW. Also, the signals sent through these
two antennas, denoted by x1 and x2 , should also be correlated with power
E[x1 x2 ] = (0.25 0.01i) mW and E[x2 x1 ] = (0.25 + 0.01i) mW.
It can be observed from Tables 3.3 and 3.4 that not every node allo-
cates its full power and total bandwidth for its outgoing links. For exam-
ple, N14 has only one outgoing link (14, 7). We can see that W(14,7) = 14.2
MHz and Tr(Q(14,7) ) = 79.47 mW, which shows N14 does not transmit at
full power and does not utilize all of its assigned bandwidth. Such a link
is not a bottleneck link. Even if the transmitter of a nonbottleneck link
increases its total transmit power and bandwidth, the end-to-end session rate
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
N14
1000 125
18.4 N5
N7
62.8 18.4
800 N6
N10 24.6
19.2 N11
15.2 N9
30.2 12.8
N13
43.4
(m)
600 47.6 9
1.4 N15
N12
38.6 N3 N2
N8 11.4
113.6
400
N4 125
200 N1
0
0 200 400 600 800 1000 1200
(m)
N14
1000
23.2 N5
N7 153.6
65.6 N6
800 23.2 N9
N10 99.2
51.4 N11 37
91.8
46.2 70.8 104.8 48.8
N13
17.8 N3
(m)
200 N1
0
0 200 400 600 800 1000 1200
(m)
cannot be increased due to the minimum bottleneck link along the path.
In this network, it can be verified that N3, N7, N11, and N12 are bottle-
neck nodes. For example, at N11, W(11,3) + W(11,5) + W(11,7) + W(11,12) = 20
MHz and Tr(Q(11,3) ) + Tr(Q(11,5) ) + Tr(Q(11,7) ) + Tr(Q(11,12) ) = 100 mW.
This means that the power and bandwidth at N11 has been fully utilized and
cannot be further increased.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
N14
1000
69.4 N5
N7
800 0.6 188.6 N6
N10 32.4
36.4
0.6
N11 N9
154.6 70.4
N13
33
600 N15
(m)
N12
N3 N2
N8 70.4
187.6
400
N4
200 N1
0
0 200 400 600 800 1000 1200
(m)
55
45
40
0 20 40 60 80 100 120 140 160
Number of iterations
3.9 Problems
3.1 Which of the following sets are convex and which are not? Why?
(a) The set in Fig. 3.6.
(b) The set in Fig. 3.7.
(c) {(x1 , x2 ) : x12 + x22 4}.
(d) {(x1 , x2 ) : x1 2x22 0}.
(e) {(x1 , x2 ) : 1 x12 + x22 4}.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
59 3.9 Problems
link 2 link 3
N2
3.2 Which of the following functions are convex, concave, affine, or neither?
Why?
(a) f (x) = x 2 + x 4 , x R.
(b) f (x) = x + x 3 , x R.
(c) f (x) = B log2 (1 + x), B is a positive constant and x R+ .
(d) f (x1 , x2 ) = x1 x2 , for x1 , x2 R+ .
(e) f (x1 , x2 ) = 5x1 10x2 , for x1 , x2 R.
3.3 Compared to a nonconvex program, why is it generally easier to obtain
an optimal solution for a convex programming problem?
3.4 Consider a small network shown in Fig. 3.8. Suppose there are two ses-
sions in this network. The first session is from N1 to N3 and includes two
routing paths: one from N1 to N3 directly and the other from N1 to N3 via
N2. The second session is from N1 to N2 and includes the direct routing path
only. Suppose y1(1) = f1 , y2(1) = y3(1) = f2 , y2(2) = f3 . Give the expressions for
A (the node-link incidence matrix), Y (the collection of all flow vectors), and
R (the collection of all source-sink vectors) in the problem formulation.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
3.5 In the networking research community, why is the sum of the log func-
tion of session rates a preferred objective function than the sum of rates?
3.6 In the formulation of Problem P in this chapter, which constraints are
associated with the transport, network, link, and physical layers, respectively?
Which constraints are associated with multiple layers?
3.7 The Problem P discussed in this chapter is a convex program. Instead of
solving the problem directly using standard convex optimization approaches,
in this chapter we solve the Lagrangian dual problem and then recover an opti-
mal solution to the original problem. Why is this approach appealing to the
networking research community?
3.8 In the dual problem decomposition, different Lagrangian dual prob-
lems can be devised for the given Problem P, depending on which con-
straints are treated as g(x) 0 and how set X is defined. Why is the constraint
< 1, YT el > c(Wl , Ql ) 0 chosen as g(x) 0?
3.9 In this chapter, a cutting-plane method is employed to solve the
Lagrangian dual problem. Explain the basic idea of this method.
3.10 In Algorithm 3.1, the third step says if zk (uk ), then (z(k) , u(k) ) is
an optimal dual solution. Why?
3.11 Solve the following problem via a dual-based approach:
Minimize x12 ln(x2 )
subject to x1 + 2x2 5 0
x1 , x2 0.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CHAPTER
When one door of happiness closes, another opens; but often we look so long at
the closed door that we do not see the one which has opened for us.
Helen Keller
61
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Wireless channels are considered unreliable due to signal fading. Spatial diver-
sity, in the form of employing multiple antennas (i.e., MIMO), has proved
to be very effective in increasing network capacity and reliability. However,
equipping a wireless node with multiple antennas may not always be practical,
as the footprint of multiple antennas may not fit on a wireless node (particularly
a handheld device). In order to achieve spatial diversity without requiring mul-
tiple antennas on the same node, the so-called cooperative communications
(CC) has been introduced [87; 136; 137]. Under CC, each node is equipped
with only a single antenna, and spatial diversity is achieved by exploiting
the antennas on other nodes in the network through cooperative relaying. In
Section 4.3, we give a brief overview of CC so as to set the stage of our case
study.
As expected, the choice of a relay node plays a critical role in the perfor-
mance of CC [18; 22; 180]. An improperly chosen relay node may offer a
smaller data rate for a sourcedestination pair than that under direct trans-
mission. In Section 4.4, we describe the relay node assignment problem in a
network environment, where multiple source-destination pairs compete for the
same pool of relay nodes in the network. Our objective is to assign the avail-
able relay nodes to different sourcedestination pairs so as to maximize the
minimum data rate among all pairs.
In Section 4.5, we will show that this problem can be formulated as a mixed-
integer linear programming (MILP). In general, an MILP has a nonpolyno-
mial complexity. In Section 4.6, instead of studying this MILP, we develop an
optimal polynomial-time algorithm, called ORA, that directly solves the orig-
inal problem. A novel idea in this algorithm is a linear marking mechanism,
which yields a low complexity for each iteration. In Section 4.7, we give a
proof of the optimality of ORA and Section 4.8 presents numerical results.
Section 4.9 summarizes this chapter. The case study in this chapter provides
an interesting instance to show that a complex problem formulation (seemingly
corresponding to a nonpolynomial-time solution) cannot be used as a way to
show that the problem is not in P.
r r r
s d s d s d
(a) One source, one destina- (b) Source node transmits in (c) Relay node transmits in the
tion, and one relay node. the first time slot. second time slot.
Figure 4.1 A three-node relay r is a relay node. Transmission from s to d is done on a frame-by-frame basis.
channel for CC. Within a frame, there are two time slots. In the first time slot, source node s
makes a transmission to the destination node d. Due to the broadcast nature
of wireless communications, this transmission is also overheard by the relay
node r. In the second time slot, node r forwards the data received in the first
time slot to node d. Note that such a two-slot structure is necessary for CC due
to the half-duplex nature of most wireless transceivers.
In this section, we provide expressions for the achievable data rate under CC
and direct transmissions (i.e., no cooperation). For CC, we consider both AF
and DF coding schemes at the relay node [87].
Amplify-and-forward (AF) Let hsd , hsr , and hrd capture the effect of
path-loss, shadowing, and fading between nodes s and d, s and r, and r and
d, respectively. Denote zd [t] and zr [t] as the zero-mean background noise at
nodes d and r in the tth time slot, respectively, with variance d2 and r2 .
Denote xs as the signal transmitted by source node s in the first time slot.
Then the received signal at destination node d, ysd , can be expressed as
ysd = hsd xs + zd [1], (4.1)
and the received signal at the relay node r, ysr , is
ysr = hsr xs + zr [1]. (4.2)
In the second time slot, relay node r transmits to destination node d. The
received signal at d, yrd , can be expressed as
yrd = hrd r ysr + zd [2],
where r is the amplifying factor at relay node r and ysr is given in (4.2). Thus,
we have
yrd = hrd r (hsr xs + zr [1]) + zd [2]. (4.3)
The amplifying factor r at relay node r should satisfy the power constraint
r2 (|hsr |2 Ps + r2 ) = Pr , where Ps and Pr are the transmission powers at
nodes s and r, respectively. So, r is given by
Pr
r2 = .
|hsr |2 Ps + r2
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
We can rewrite (4.1), (4.2), and (4.3) in the following compact matrix form:
Y = Hxs + BZ,
where
zr [1]
ysd hsd 0 10
Y= , H= , B= , and Z = zd [1] .
yrd r hrd hsr r hrd 01
zd [2]
(4.4)
It has been shown in [87] that the above channel, which combines both the
direct path (s to d) and the relay path (s to r to d), can be modeled as a
one-input, two-output complex Gaussian noise channel. The achievable data
rate CAF (s, r, d) from s to d is given by
W ! "
CAF (s, r, d) = log2 det I + (Ps HH )(BE[ZZ ]B )1 , (4.5)
2
where W is the bandwidth, det() is the determinant function, I is the identity
matrix, the superscript represents the complex conjugate transpose, and
E[] is the expectation function.
Substituting (4.4) into (4.5) and performing algebraic manipulations, we
have
# $
W Ps Ps |hsr |2 Pr |hrd |2
CAF (s, r, d) = log2 1 + 2 |hsd | +
2
.
2 d Ps d2 |hsr |2 + Pr r2 |hrd |2 + r2 d2
(4.6)
Denote
Ps Ps Pr
SNRsd = |hsd |2 , SNRsr = 2 |hsr |2 , and SNRrd = 2 |hrd |2 .
d2 r d
We have
where
% &
1 SNRsr SNRrd
IAF (SNRsd , SNRsr , SNRrd ) = log2 1 + SNRsd + .
2 SNRsr + SNRrd + 1
where
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Based on the above results, we have two observations. First, comparing CAF
(or CDF ) to CD , we can see that CC may not always be better than direct trans-
missions. For example, a poor choice of relay node could make the achievable
data rate under CC lower than that under direct transmissions. This fact under-
lines the significance of relay node selection in CC. Second, although AF and
DF are different mechanisms, the data rates for both of them have the same
form, i.e., a function of SNRsd , SNRsr , and SNRrd . Therefore, we only need
to develop one optimal relay node assignment algorithm, which should work
for both AF and DF. Table 4.1 lists the notations used in this chapter.
Based on the background in the last section, we consider a relay node assign-
ment problem in a network setting. There are N nodes in an ad hoc network,
with each node being either a source node, a destination node, or a poten-
tial relay node (see Fig. 4.2). In order to avoid interference, we assume that
orthogonal channels are available in the network (e.g., using OFDMA) for
CC [87]. The path-loss between nodes u and v is captured in huv and is
given a priori. Denote Ns = {s1 , s2 , . . . , sNs } as the set of source nodes, Nd =
{d1 , d2 , . . . , dNd } as the set of destination nodes, and Nr = {r1 , r2 , . . . , rNr } as
the set of relays (see Fig. 4.2). We consider unicast where every source node
si is paired with a destination node di , i.e., Nd = Ns . Each node is equipped
with a single transceiver and can transmit/receive within one channel at a time.
Further, we assume that each node can only serve a unique role of source,
destination, or relay. That is, Nr = N 2Ns .
Note that a source node may not always get a relay node. There are two
possible scenarios in which this may happen. First, there may not be a sufficient
number of relay nodes in the network (e.g., Nr < Ns ). In this case, some source
nodes will not have relay nodes. Second, even if there are enough relay nodes,
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Symbol Definition
a sender may choose not to use a relay node if it leads to a lower data rate than
direct transmission.
We now discuss the objective function of our problem. Although different
objectives can be used, a widely used objective for CC is the achievable data
rate. For the multi-session network environment considered in this chapter
(see Fig. 4.2), each sourcedestination pair will have a different achievable
data rate after we apply a relay node assignment algorithm. So a plausible
objective is to maximize the minimum achievable data rate among all the
sourcedestination pairs.
More formally, denote R(si ) as the relay node assigned to si , and S(rj ) as
the source node that uses rj . For both AF and DF, its achievable data rate can
be written as (see Section 4.3)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CR (si , ) = CD (si , di ).
Note that di is not listed in function CR (si , R(si )) since for each source node
si , its corresponding destination node di is deterministic.
Denote by Cmin the objective function, which is the minimum achievable
data rate among all source nodes. We have
Our goal is to find an optimal relay node assignment for all the source
destination pairs such that Cmin is maximized.
In this section, we present a formulation for the relay node assignment problem
based on an optimization approach from OR. To describe relay node assign-
ment mathematically, we denote a binary variable Asi ,rj as
'
1 if relay rj is assigned to source si ,
Asi ,rj =
0 otherwise.
Since a source node si will be assigned at most one relay node, we have
Nr
Asi ,rj 1.
j =1
Due to the unique role of each node, a relay node rj can be assigned to at most
one source node. That is
Ns
Asi ,rj 1.
i=1
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Maximize Cmin
Nr
subject to Cmin 1 Asi ,rj CR (si , )
j =1
Nr
+ Asi ,rj CR (si , rj ) (si Ns )
j =1
Ns
Asi ,rj 1 (rj Nr )
i=1
Nr
Asi ,rj 1 (si Ns )
j =1
Asi ,rj {0, 1}, Cmin 0 (si Ns , rj Nr )
Note that if any Asi ,rj = 1, then a relay rj is assigned to source si . By (4.7),
the achievable data rate from source si to destination di by using relay rj
r
is CR (si , rj ). On the other hand, if N j =1 Asi ,rj = 0, then there is no relay
assigned to source si . By (4.7), the achievable data rate from source si to des-
tination di is CR (si , ). Combining these two cases, the achievable data rate
from source si to destination di is
Nr
Nr
1 Asi ,rj CR (si , ) + Asi ,rj CR (si , rj ).
j =1 j =1
Based on the above discussion, one formulation for the optimal relay node
assignment problem is given in Table 4.2. This formulation is a mixed-integer
linear programming (MILP) and is NP-hard in general [46]. As discussed in
Section 4.1, this does not mean that our optimal relay node assignment prob-
lem is NP-hard because an NP-hard formulation is not a proof of a problems
NP-hardness. In fact, this problem can be viewed as a maximin linear assign-
ment problem, for which there exists polynomial-time algorithms (see [3],
[25], [89], [126], and [161] for example, although these papers address dif-
ferently structured problems, the proposed algorithms can be adapted to solve
our problem in polynomial time as well).
Instead of applying a general algorithm for the maximin linear assignment
problem, we design in the next section a specialized polynomial-time algo-
rithm (along with an optimality proof) for our relay node assignment problem.
The existence of such a polynomial-time algorithm implies that the problem
belongs to the class P of polynomially solvable problems.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
YES
Can we find
a better solution?
Use Find_Another_Relay(sb) to
improve the data rate of sb,
NO and return the outcome.
END
Find_Another_Relay(si )
BEGIN
Can we find an
Return NO unmarked relay for si
NO with data rate larger
than Cmin?
YES
NO Is this relay
already assigned?
YES
For sj , use
YES Find_Another_Relay(sj) to NO
Better solution found. determine if another
relay can be
assigned
Clear marks on
all relays.
Return
YES
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
r3 s3 r5 s5 r6 s6
can be
assigned to s6
can offer an increase in its data rate compared to direct transmissions, i.e.,
those relays with CR (si , rj ) > CR (si , ). Obviously, it only makes sense to
consider these relays for CC. That is, for a source node si , those relay nodes
with rates not greater than direct transmissions can be removed from further
consideration.
After the preprocessing step, we enter the initial assignment step. The objec-
tive of this step is to obtain an initial feasible solution for the ORA algorithm
so that it can start its iteration. In the preprocessing step, we have already iden-
tified the list of relay nodes for each source node that can increase its data rate
compared to direct transmission. We can randomly assign a relay node from
this list to a source node. Note that once a relay node is assigned to a source
node, it cannot be assigned again to another source node. If there is no relay
node available for a source node, then this source node will simply employ
direct transmission. Upon the completion of this assignment, each source node
will have a data rate that is no less than that under direct transmission.
The next step in the ORA algorithm is to find a better assignment, which
is an iterative process. This is the key step in the ORA algorithm. The detail
of this step is shown in the bottom portion of Fig. 4.3. As a starting point of
this step, the ORA algorithm identifies the smallest data rate Cmin among all
sources. The ORA algorithm aims to increase this minimum data rate for the
corresponding source node, while having all other source nodes maintain their
data rates above Cmin . Without loss of generality, we use Fig. 4.4 to illustrate
the search process.
Suppose ORA identifies that s1 has the smallest rate Cmin under the current
assignment (with relay node r1 ). In case of a tie, i.e., when two or more
source nodes have the same smallest data rate, the tie is broken by choosing
the source node with the highest node index. Source s1 examines other relays
with a rate larger than Cmin . If it cannot find such a relay, then no better
solution is found and the ORA algorithm terminates.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Otherwise, i.e., when there exist better relays, we consider these relays in a
non-increasing order of data rate (should it be assigned to s1 ). That is, we try
the relay that can offer the maximum possible increase in data rate first. In
case of a tie, i.e., when two or more relay nodes offer the same maximum data
rate, the tie is broken by choosing the relay node with the highest node index.
Suppose that source node s1 considers relay node r2 . If this relay node is not
yet assigned to any other source node, then r2 can be immediately assigned
to s1 . In this simple case, we find a better solution and the current iteration
is completed.
Otherwise, i.e., if r2 is already assigned to a source node, say s2 , we mark
r2 to indicate that r2 is under consideration and check whether r2 can be
released by s2 .
To release r2 , source node s2 needs to find another relay (or use direct trans-
mission) while making sure that such a new assignment still keeps its data
rate larger than the current Cmin . This process is identical to that for s1 , with
the only (but important) difference being that s2 will not consider a relay
that has already been marked, because that relay node has already been
considered by a source node encountered earlier in this iteration.
Suppose that source node s2 now considers relay r3 . We consider the follow-
ing three cases. If this relay node is not yet assigned to any source node, then
r3 can be assigned to s2 , r2 can be assigned to s1 , and the current iteration is
completed. Moreover, if the relay under consideration by s2 is the one that is
being used by the source node that initiated the iteration, i.e., relay r1 , then
it is easy to see that r1 can be taken away from s1 . A better solution (r1 is
assigned to s2 and r2 is assigned to s1 ) is found and the current iteration is
completed. Otherwise, we mark r3 and check further to see whether r3 can
be released by its corresponding source node, say s3 . We also note that s2
can consider direct transmission if it offers a data rate larger than Cmin .
Suppose that s3 cannot find any unmarked relay that offers a data rate
larger than Cmin and its achievable data rate under direct transmission is not
larger than Cmin . Then s2 cannot use r3 as its relay.
If any unmarked relay that offers a data rate larger than Cmin cannot be
assigned to s2 , then s1 cannot use r2 and the algorithm will move on to con-
sider the next relay on its nonincreasing rate list, say r4 .
The search continues, with relay nodes being marked along the way, until a
better solution is found or no better solution can be found. For example,
in Fig. 4.4, s6 finds a new relay r7 . As a result, we have a new assign-
ment, where r7 is assigned to s6 , r6 is assigned to s4 , and r4 is assigned
to s1 .
Note that the mark on a relay node will not be cleared throughout the
search process in the same iteration. We call this a linear marking mech-
anism. These marks will only be cleared when the current iteration termi-
nates and before the start of the next iteration. A pseudocode for the ORA
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
r1 r2 r3 r4 r5 r6 r7
s1 14 7 24 5 14 15 17 9
s2 9 8 10 11 20 10 12 11
s3 11 10 13 17 21 8 9 19
s4 12 8 9 12 11 10 9 18
s5 10 9 18 19 24 9 13 23
s6 7 18 12 6 11 11 17 20
s7 16 1 9 4 14 19 8 12
(b) The start of the second iteration.
s1 14 7 24 5 14 15 17 9
s2 9 8 10 11 20 10 12 11
s3 11 10 13 17 21 8 9 19
s4 12 8 9 12 11 10 9 18
s5 10 9 18 19 24 9 13 23
s6 7 18 12 6 11 11 17 20
s7 16 1 9 4 14 19 8 12
(c) The start of the third iteration.
s1 14 7 24 5 14 15 17 9
s2 9 8 10 11 20 10 12 11
s3 11 10 13 17 21 8 9 19
s4 12 8 9 12 11 10 9 18
s5 10 9 18 19 24 9 13 23
s6 7 18 12 6 11 11 17 20
s7 16 1 9 4 14 19 8 12
(d) Final assignment upon ORA termination.
s1 14 7 24 5 14 15 17 9
s2 9 8 10 11 20 10 12 11
s3 11 10 13 17 21 8 9 19
s4 12 8 9 12 11 10 9 18
s5 10 9 18 19 24 9 13 23
s6 7 18 12 6 11 11 17 20
s7 16 1 9 4 14 19 8 12
Example 4.1
Suppose that there are seven sourcedestination pairs and seven relay nodes
in the network.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 4.3(a) shows the data rate for each source node si when relay node
rj is assigned to it. The symbol indicates direct transmission. Also shown
in Table 4.3(a) is an initial relay node assignment, which is indicated by an
underscore on the intersecting row (si ) and column (rj ). That is, the initial
assignment is r5 for s1 , r4 for s2 , r2 for s3 , r7 for s4 , r3 for s5 , r6 for s6 , and
for s7 . Note that the initial assignment step ensures that the achievable
data rate for each sourcedestination pair is no less than that under direct
transmission.
Given the initial relay node assignment in Table 4.3(a), source s3 is iden-
tified as the bottleneck source node sb , with the current Cmin = 13. Since
consideration of relay nodes is given in the order of nonincreasing (from
largest to smallest) data rate, r4 is therefore considered for s3 first. Since r4
is already assigned to source node s2 , we mark r4 now. Next we check
whether or not s2 can find another relay. But any other relay (or direct trans-
mission) will result in a data rate no greater than the current Cmin = 13. This
means that r4 cannot be taken away from s2 . Therefore, s3 will then con-
sider the next relay node that offers the second largest rate value, i.e., r7 .
Since r7 is already assigned to sender s4 , we mark r7 . Next, the ORA
algorithm will check whether or not s4 can find another relay. But none of
the relay nodes except r7 can offer a data rate larger than the current Cmin to
s4 . So r7 cannot be taken away from s4 . Therefore, node s3 will now check
for the relay node that offers the next highest rate, i.e., r3 . Since r3 is already
assigned to s5 , we mark r3 now. Next, the ORA algorithm checks to see
if s5 can find another relay; s5 checks relay nodes in nonincreasing order of
data rate values. Since r4 (with largest rate) and r7 (with the second largest
rate) are both marked, they will not be considered (this is the essence of
the linear marking mechanism of the ORA algorithm, which ensures that
the linear complexity of examining relay nodes in each iteration). The next
relay to be considered is r2 , which offers a rate of 18 > Cmin = 13. More-
over, r2 is the relay node assigned to sb = s3 . Thus, r2 can be reassigned to
s5 and r3 can be reassigned to s3 . The new assignment after the first itera-
tion is shown in Table 4.3(b). Now the objective value, Cmin , is updated to
15, which corresponds to s1 . Before the second iteration, all markings done
in the first iteration are cleared.
In the second iteration, the ORA algorithm will identify s1 as the source
node with the minimum data rate in the network. The algorithm will then
perform a new search for a better relay node for source s1 . Similar to the
first iteration, the assignments for other source nodes can change during
this search process, but all assignments should result in data rates larger
than 15. The details of this iteration are similar to those in the first iteration
and we leave them as a homework exercise.
Similar to the previous two iterations, in the third iteration, the ORA
algorithm will identify the source node with minimum data rate, which is
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
now s7 . The algorithm will then perform a new search to improve the data
rate of s7 .
Finally, in the fourth and final iteration, the algorithm will try to improve
the data rate of s3 and will be unsuccessful. As a result, the ORA algorithm
will terminate.
The final assignment upon termination of the ORA algorithm is shown
in Table 4.3(d), with the optimal (maximum) value of Cmin being 17.
throughout the iteration. For example, in Fig. 4.4, when s4 tries to find another
relay, it will no longer consider r2 and r3 that have already been marked.
Similarly, when s6 tries to find another relay, it will not consider r2 , r3 , r4 , and
r5 . As a result, any relay node will be considered at most once in the search
process and the pool of candidate relay nodes shrinks linearly as the search
continues. It is not hard to see that this marking mechanism leads to a linear
complexity of examining relay nodes in each iteration of the ORA algorithm.
Although such linear marking mechanism is attractive in its complexity, a
natural question we would ask is: How could such a linear marking lead to
an optimal solution, as it may overlook many possible assignments that poten-
tially increase Cmin ? This is precisely the question that we will address in
Section 4.7, where we will prove that ORA can guarantee that its final solution
is optimal (Theorem 4.1).
Remark 4.1
We comment here that the linear marking mechanism can be alternatively
viewed as a flow augmentation step within a specially defined max-flow
problem (see Problem 4.12 in the exercise).
In this section, we give a correctness proof of the ORA algorithm. That is,
upon termination of the ORA algorithm, the solution (i.e., objective value and
the corresponding relay node assignment) is optimal.
Our proof is based on contradiction. Denote as the final solution obtained
by the ORA algorithm, with the objective value being Cmin . For , denote the
relay node assigned to source node si as R (si ). Conversely, for , denote the
source node that uses relay node rj as S (rj ).
Assume that there exists a better solution than . That is, the objective
value for , denoted as Cmin , is greater than that for , i.e., Cmin > Cmin . For
, denote the relay node assigned to source node si as R (si ), the source node
that uses relay node rj as S (rj ).
The key idea in the proof is to exploit the marking status of the ORA algo-
rithm at the end of the last iteration, which is a nonimproving iteration. Specif-
ically, in the beginning of this last iteration, ORA will select a bottleneck
source node, which we denote as sb . ORA will then try to improve the solution
by searching for a better relay node for this bottleneck source node. Since the
last iteration is a nonimproving iteration, ORA will not find a better solution,
and thus will terminate. We will show that R (sb ) is not marked at the end of
the last iteration of ORA. On the other hand, if there exists a better solution
than , we will show that R (sb ) is marked at the end of the last iteration of
ORA. This leads to a contradiction and thus the assumption of the existence of
a better solution cannot hold. We begin our proof with the following fact:
Fact 4.1
For the bottleneck source node sb under , its relay node R (sb ) is not
marked at the end of the last iteration of the ORA algorithm.
^ R^( Gk(sb) )
Gk(sb)
(marked)
^ R^( Gn(sb) )
Gn(sb)
(marked)
Claim 4.1
Relay node R (sb ) must be marked at the end of the last iteration of the
ORA algorithm. Further, it cannot be and must be assigned to some source
node under solution .
Proof. Since is a better solution than , we have CR (sb , R (sb )) Cmin >
Cmin . Thus, by construction, ORA will consider the availability of relay node
R (sb ) for sb in its last iteration of . Since the ORA algorithm cannot find a
better solution in this last iteration, R (sb ) must be marked and then the out-
come of checking the availability of R (sb ) shows that it is unavailable. By
the linear marking process, the mark on R (sb ) will not be cleared through-
out the search process in the last iteration. Thus, R (sb ) is marked at the end
of the last iteration of the ORA algorithm.
We now prove the second statement by contradiction. If R (sb ) were
, then sb would choose in the last iteration of ORA since it can offer
CR (sb , R (sb )) > Cmin . But this contradicts the fact that we are now in the
last iteration of ORA, which is a nonimproving iteration. So R (sb ) cannot be
. Further, since we proved that R (sb ) is marked at the end of the last iteration
of the ORA algorithm, it must be assigned to some source node already.
G () = S (R ()). (4.8)
Thus, relay node R (sb ) is assigned to source node G (sb ) in (see Fig. 4.5).
Since R (sb ) = R (sb ), they are assigned to different source nodes in ,
i.e., G (sb ) = sb . Now we recursively investigate the relay node assigned to
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
G (sb ) under the solution , i.e., R (G (sb )). We have the following claim
(also see Fig. 4.5):
Claim 4.2
Relay node R (G (sb )) must be marked at the end of the last iteration of
the ORA algorithm. Further, it cannot be and must be assigned to some
source node under the solution .
The proofs for both statements in this claim follow the same token as that
for Claim 4.1 and are thus left as a homework exercise.
Again, by the definition of S (), we have that the relay node R (G (sb ))
is assigned to source node S (R (G (sb ))) in solution . By (4.8), we have
source node S (R (G (sb ))) = G (G (sb )). To simplify the notation, we
define the function G2 () as
G2 () = G (G ()).
Claim 4.3
Relay node R (Gk (sb )) must be marked at the end of the last iteration of
the ORA algorithm. Further, it cannot be and must be assigned to some
source node under solution , k = 0, 1, 2, . . . , n.
solution obtained by ORA, whereas R (Gk (sb )) is the relay node assigned
to this source node in the hypothetical better solution . Our goal is to show
that ORA should have marked this relay node in its last iteration.
Since CR (Gk (sb ), R (Gk (sb ))) > Cmin and R (Gk (sb )) is not assigned to
G (sb ) in the last iteration of ORA, then by construction of ORA, ORA must
k
have checked the availability of R (Gk (sb )) for Gk (sb ) during the last iter-
ation, then marked it, and then determined that it is unavailable for Gk (sb ).
Moreover, due to the linear marking process, this mark on R (Gk (sb ))
should remain after the last iteration of ORA. Thus, we conclude that
R (Gk (sb )) is marked at the end of the last iteration of the ORA algorithm.
We now prove the second statement by contradiction. If R (Gk (sb ))
is , then Gk (sb ) will choose in the last iteration since it can offer
CR (Gk (sb ), R (Gk (sb )) > Cmin , and finally sb will be able to get a better
relay node. But this contradicts the fact that this last iteration is a nonim-
proving iteration. So, R (Gk (sb )) cannot be . Further, since we proved that
R (Gk (sb )) is marked at the end of the last iteration of the ORA algorithm,
it must be assigned to some source node already.
Referring to Fig. 4.5, we have Claim 4.3 for a set of relay nodes
R (sb ), R (G (sb )), . . . , R (Gn (sb )). Our recursive procedure terminates
at R (Gn (sb )) because its assigned source node in solution is a node in
{sb , G (sb ), . . . , Gn (sb )}. We are now ready to prove the following theorem,
which is the main result of this section:
Theorem 4.1
Upon termination of the ORA algorithm, the obtained solution is optimal.
Proof. Suppose that the solution obtained by the ORA algorithm is not opti-
mal. Then, there exists a better solution . We have Fact 4.1 and Claim 4.3.
We will show a contradiction.
Under Claim 4.3, we proved that the relay node R (Gn (sb )) is assigned to a
source node in the solution obtained by ORA. Since our recursive procedure
terminates at R (Gn (sb )), its assigned source node in solution is a node
in {sb , G (sb ), . . . , Gn (sb )}. But we also know that under , the source nodes
G (sb ), G2 (sb ), G3 (sb ), . . . , Gn (sb ) have relay nodes R (sb ), R (G (sb )),
R (G2 (sb )), . . . , R (Gn1 (sb )), respectively. Thus, R (Gn (sb )) can only be
assigned to sb in solution . On the other hand, the relay node assigned to sb in
the solution is denoted by R (sb ). Therefore, we must have R (Gn (sb )) =
R (sb ).
Now, Claim 4.3 states that R (Gn (sb )) must be marked after the last
iteration, whereas Fact 4.1 states that the relay node assigned to the bottle-
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
neck source node, i.e., R (sb ), cannot be marked. Since both R (sb ) and
R (Gn (sb )) are the same relay node, we have a contradiction. Thus, our
assumption that there exists a better solution than does not hold.
Note that the proof of Theorem 4.1 does not depend on the initial assign-
ment in the ORA algorithm. So we have the following corollary for the ORA
algorithm:
Corollary 4.1
Under any feasible initial relay node assignment, the ORA algorithm can
find an optimal relay node assignment.
4.8.2 Results
Case 1: Nr Ns . In this case (see Fig. 4.6), we have 30 sourcedestination
pairs and 40 relay nodes.
Under ORA, after preprocessing, we start with an initial relay node assign-
ment in the first iteration. Such an initial assignment is not unique. But regard-
less of the initial relay node assignment, we expect the objective value Cmin
to converge to the optimum (Corollary 4.1). To validate this result, we show
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 4.4 Locations and roles of all the nodes in the network.
the results of running the ORA algorithm under two different initial relay node
assignments, denoted as I and II (see Table 4.5).
In Table 4.5, the second column shows the data rate for each source
destination pair under direct transmissions. Note that the minimum rate among
all pairs is 1.83 Mbps, which is associated with s7 . The third to fifth columns
are results under the initial relay node assignment I and the sixth to eighth
columns are results under the initial relay node assignment II. The symbol
denotes direct transmission. Note that initial relay node assignments I and
II are different. As a result, the final assignment is different under I and II.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
However, the final objective value (i.e., Cmin ) under I and II is identical (4.43
Mbps).
Fig. 4.7 shows the objective value Cmin at each iteration under initial relay
node assignments I and II. Under either initial assignments, Cmin is a nonde-
creasing function of iteration number. Note that a higher initial value of Cmin
does not mean that ORA will converge faster.
Case 2: Nr < Ns . In this case (see Fig. 4.8), we have 40 sourcedestination
pairs and 20 relay nodes.
Table 4.6 shows the results of this case under two different initial relay node
assignments I and II. The second column in Table 4.6 lists the data rate under
direct transmissions for each sourcedestination pair. As discussed at the end
of Section 4.6.2, for the case of Nr < Ns , it is only necessary to consider relay
node assignments for Nr = 20 source nodes corresponding to the 20 small-
est data rates, i.e., nodes s1 , s3 , s4 , s5 , s7 , s8 , s11 , s12 , s13 , s14 , s15 , s16 , s17 , s21 ,
s24 , s25 , s27 , s28 , and s29 .
Again in Table 4.6, the objective value Cmin is identical (3.80 Mbps) regard-
less of different initial relay node assignments (I and II). Note that despite the
different final relay node assignments under I and II, the objective value Cmin
is identical.
Fig. 4.9 shows the objective value Cmin at each iteration under initial relay
node assignments I and II. Again, we observe that in Fig. 4.9, Cmin is a nonde-
creasing function of iteration number under both initial relay node assignments
I and II.
Significance of preprocessing Now we use a set of numerical results to
show the significance of preprocessing in our ORA algorithm. We consider
the same network in Fig. 4.6 with 30 sourcedestination pairs and 40 relay
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 4.5 Optimal assignments for Case 1 (Nr Ns ) under two different initial relay
node assignments.
0
0 100 200 300 400 500 600 700 800 900 1000 1100 12001300 1400 1500 1600
(meters)
On the other hand, when the preprocessing step is employed, ORA can
ensure that the final rate for each sourcedestination pair is no less than that
under direct transmission, as shown in Table 4.5.
Table 4.6 Optimal assignments for Case 2 (Nr Ns ) under two different initial relay
node assignments.
Without preprocessing
Sender CD (Mbps) Initial Final Final rate (Mbps)
s1 2.62 r3 r3 6.54
s2 4.60 4.60
s3 3.81 r2 8.73
s4 2.75 r8 r4 4.66
s5 3.15 r14 r14 6.47
s6 4.17 r6 9.25
s7 1.83 r6 r8 4.76
s8 2.99 r12 7.22
s9 4.92 4.92
s10 4.80 4.80
s11 4.13 r20 9.13
s12 3.24 r18 5.55
s13 3.68 r17 7.32
s14 4.23 r16 7.87
s15 2.62 r19 5.84
s16 3.30 r22 7.30
s17 4.17 r24 5.62
s18 6.03 r23 r21 7.37
s19 8.76 r39 r39 4.81
s20 6.95 r26 r26 7.25
s21 1.90 r27 4.90
s22 7.65 r28 r28 8.71
s23 7.55 r29 r29 11.26
s24 2.12 r40 4.43
s25 3.91 r30 5.87
s26 6.08 r33 r33 7.55
s27 3.61 r34 5.45
s28 2.04 r35 5.29
s29 2.33 r31 4.68
s30 6.60 6.60
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
89 4.10 Problems
4.10 Problems
4.1 Can a generally NP-hard formulation for a problem be used to claim that
the problem is NP-hard? Why?
4.2 Verify Eq. (4.6) in this chapter.
4.3 Show that IAF (SNRsd , SNRsr , SNRrd ) is an increasing function of Ps
and Pr , respectively. Do the same for IDF (SNRsd , SNRsr , SNRrd ).
4.4 Consider the simple three-node relay channel in Fig. 4.1. The locations
of the source node s and the destination node d are at (0, 0) and (100, 0),
respectively, all in meters. The maximum transmit power of both source and
relay nodes is 0.1 W. The channel coefficient between two nodes is h = 2d 2 ,
where d is the distance between the two nodes. The noise power spectrum
density (PSD) at all nodes is 2 1016 W/Hz and the channel bandwidth is 10
MHz.
(a) Calculate the achievable rate under direct transmission between s and d.
(b) If the location of the relay node r is at (50, 33) (in meters), calculate the
achievable rates under AF and DF, respectively.
(c) If the location of the relay node r is at (80, 50) (in meter), calculate the
achievable rates under AF and DF, respectively.
4.5 If DF is employed for CC, how will it affect the operation of ORA?
Justify your answer.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
r1 r2 r3 r4 r5 r6
s1 4 17 8 15 5 8 15
s2 6 12 8 16 14 8 13
s3 3 15 6 21 13 14 13
s4 11 19 15 12 16 9 15
r1 r2 r3 r4 r5 r6 r7
s1 1 17 6 15 10 20 11 15
s2 6 11 6 17 9 18 10 13
s3 6 5 18 20 16 18 15 12
s4 4 18 12 9 6 15 19 11
s5 15 19 4 15 16 19 8 15
s6 9 6 11 16 15 18 11 15
s7 6 11 19 5 7 9 18 15
4.6 In practice, the number of relay nodes may or may not be less than the
number of source nodes, i.e., Nr < Ns or Nr Ns . How will this affect the
operation of the ORA algorithm?
4.7 Describe the marking mechanism in the ORA algorithm. Why do we call
it linear? Explain its role in the algorithms complexity.
4.8 For the naive marking mechanism described in Section 4.6.3, show that
the number of checks for relay nodes within an iteration is O(Nr Nr !) if Nr <
r Nr !
Ns and O( (NrNN s +1)!
) if Nr Ns .
4.9 For Example 4.1 in Section 4.6.2, give the details of how ORA and linear
marking work in the second and third iterations, respectively.
4.10 Apply the ORA algorithm to find an optimal solution for the data given
in Table 4.8. Show your work for each iteration.
4.11 Apply the ORA algorithm to find an optimal solution for the data given
in Table 4.9. Show your work for each iteration.
4.12 Give a complete proof of Claim 4.2.
4.13 For any given relay node assignment with an objective value Cmin , we
can define a directional graph as follows (see Fig. 4.10 as an example). The set
of nodes in the graph is Ns Nr {S, D, }, where S, D, and are dummy
source, destination, and relay nodes, respectively. The graph has the following
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
91 4.10 Problems
1
1 1 r2
s2
1
1
1
S
D
1
1
1 1
rNr Ns
sNs
1
edges: (i) edges (S, si ) with unit capacity, for each si Ns ; (ii) edges (si , rj )
with unit capacity, for each si Ns , rj Nr , CR (si , rj ) > Cmin , and edges
(si , ) with unit capacity, for each si Ns , CR (si , ) > Cmin ; and (iii) edges
(rj , D) with unit capacity, for each rj Nr , and an edge (, D) with capacity
Ns . Show how such a max-flow problem for the defined graph is related to the
problem of identifying an improving solution.
4.14 In each iteration of the ORA algorithm, we try to find a better solution.
Denote C as the set of all possible values for the maximum Cmin value in an
optimal solution. Then in the worst case, ORA decreases |C| by one in an itera-
tion. We can improve the ORA algorithm as follows. In each iteration, instead
of just finding a better solution, we identify the median value Cmed for the cur-
rent set C and try to find a solution with an objective value of at least Cmed . We
may or may not find such a solution. But for either case, we can decrease |C|
to 12 |C| and thus the number of iterations can be significantly decreased. Give
a detailed description of this enhanced ORA (EORA) algorithm and analyze
its complexity. (Hint: The complexity of one iteration in EORA is higher than
that for ORA, since it may be necessary to find better relay nodes for multiple
sources in an iteration of EORA.)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
PART
CHAPTER
5 Branch-and-bound framework
and application
Vision is not enough, it must be combined with venture. It is not enough to stare
up the steps, we must step up the stairs.
Vaclav Havel
95
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Bounding step The goal of this step is to develop a lower bound and an
upper bound for a particular problem. First of all, by using some relaxation
technique, branch-and-bound obtains a relaxation for the original problem and
its solution provides a lower bound (LB) for the objective function. Such a
relaxation is usually in the form of a LP or a convex programming problem,
which we can solve in polynomial time (given that the relaxation itself is of
polynomial size). It is clear that we should make this relaxation as tight as
possible, i.e., the obtained lower bound should be as close to the optimal value
as possible. The particular approach to develop a tight and polynomial-time
solvable relaxation needs to be carefully designed by the user.
Using the relaxation solution as a starting point, branch-and-bound employs
a local search algorithm to find a feasible solution to the original problem,
which provides an upper bound (U B) for the objective function (see Fig. 5.1(a)
for an example). Again, the local search algorithm needs to be carefully
designed by the user. This computation of lower and upper bounds constitutes
the bounding step of the branch-and-bound framework.
If the obtained lower and upper bounds are within an -tolerance of each
other, i.e., LB (1 )U B, we are done, i.e., the current feasible solution is
(1 )-optimal. If the relaxation is not tight, then the lower bound LB could
be significantly lower than the upper bound U B. To close this gap, we must
attempt to produce a tighter relaxation that yields a smaller relaxation error.
This can be achieved by a branching (or called a partitioning) step.
Branching step The branching step includes partitioning problem, updating
lower and upper bounds, and/or fathoming a subproblem.
Objective
LB=LB1
Original problem 1
(a) Iteration 1.
UB3
UB=UB2
Objective
LB2
LB=LB3
Problem 2 Problem 3
(b) Iteration 2.
UB2
UB4
Objective
UB=UB5
LB4 LB5
LB=LB2
(c) Iteration 3.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Updating lower and upper bounds For each of two new problems, the
branch-and-bound procedure solves a corresponding relaxation and performs
a local search. Hence, we derive lower bounds LB2 and LB3 for Problems
2 and 3, respectively. We also possibly obtain feasible solutions that provide
upper bounds U B2 and U B3 for Problems 2 and 3, respectively. It is possible
that a new problem does not have a feasible solution or the (heuristic) local
search procedure might fail to find a feasible solution even if one exists. But
since the feasible solution 1 to Problem 1, assuming that such a solution has
been detected, must be feasible to one of the two new problems (unless some
explored subregion that contains 1 has been deleted from consideration
during the analysis of Problem 1), we have at hand a feasible solution for at
least one new problem by using the previous feasible solution if necessary.
Since the relaxations in Problems 2 and 3 are both more tightly constrained
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
than that of Problem 1, we have min{LB2 , LB3 } LB1 . The lower bound of
the original problem is updated from LB = LB1 to LB = min{LB2 , LB3 }.
The upper bound of the original problem is updated from U B = U B1 to
U B = min{U B, U B2 , U B3 }, since the best feasible solution to the original
problem is the solution with the smallest U Bi -value.1 As a result, we now
have a smaller gap between LB and U B. If LB (1 )U B, we are done.
Otherwise, we choose a problem with the minimum lower bound (Problem
3 in Fig. 5.1(b)) and perform partitioning on this problem.
Fathoming a subproblem. Note that during the branch-and-bound process,
if we find a Problem z with LBz (1 )U B (see Problem 4 in Fig. 5.1(c)),
then we have the following two possible cases:
Case 1: The global optimal solution is not in Problem z. In this case, the
removal of Problem z will not cause the loss of the optimal solution in future
iterations.
Case 2: The global optimal solution is in Problem z. In this case, the opti-
mal (feasible) solution must have an objective value greater than or equal to
LBz , which means that it is also greater than or equal to (1 )U B (since
LBz (1 )U B). Thus, the current best feasible solution with objective
value U B is already a (1 )-optimal solution. Therefore, we can still guar-
antee (1 )-optimality if we remove this problem from the list for further
consideration.
The foregoing cases underscore the principal advantage of the branch-and-
bound framework, i.e., we can remove a problem before we completely solve
it. This removal of a subproblem from consideration in the active list of
problems is called fathoming the subproblem. Eventually, when we obtain
LB (1 )U B, the branch-and-bound procedure terminates.
1 Note that we always keep the best detected (incumbent) feasible solution in the process.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
m. For transmission from node i to node j , a simple model for path attenuation
loss gij is
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Symbol Definition
m
cij Link capacity of link i j under pij m
where dij is the physical distance between nodes i and j and is the path-
loss index. In this context, we assume data transmission from node i to node
j is successful only if the received power at node j exceeds a threshold PT .
Denote the transmission range of node i under pijm as R (p m ). Then, based on
T ij
gij pij PT and (5.1), we can calculate the transmission range of this node
m
as follows:
# m
$1/
pij
m
RT (pij ) = . (5.2)
PT
Since the receiving node j must be physically within the transmission range of
node i, we have
# m $1/
pij
(C-1) dij .
PT
4
3
6
2 5
4
3
6
5
2
'
1 if node i transmits data to node j on frequency band m,
xijm =
0 otherwise.
As mentioned earlier, we consider scheduling in the frequency domain and
assume unicast communication. Thus once a band m is used by node i for
transmission to node j , this band cannot be used again by node i to transmit to
a different node. That is,
(C-3) xijm 1,
j Tim
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
where Tim is the set of nodes to which node i can transmit on band m under
full power P .
Denote by RTmax the maximum transmission range of a node when it
transmits at full power P . Then based on (5.2), we have RTmax = RT (P ) =
1/
P
PT . Thus, we have PT = (R max P
. Then, for a node transmitting at a
T )
power p [0, P ], its transmission range is given by
% &1 1 p 1
p p(RTmax )
RT (p) = = = RTmax . (5.3)
PT P P
p 1
RI (p) = RImax . (5.4)
P
Recall that Tim denotes the set of nodes to which node i can transmit on
band m under full power P . More formally, we have
Similarly, denote by Ijm the set of nodes that can interfere with node j on band
m under full power P . Then,
Ijm = {k : dj k RImax , m Mk }.
Note that the definitions of Tim and Ijm are both based on full transmission
power P . When the power level p is below P , the corresponding transmission
and interference ranges will be smaller. Nevertheless, the set of nodes that fall
in the transmission range and the set of nodes that can produce interference
can be upper bounded by those sets under full transmission power.
By (C-1) and (C-2) (i.e., constraints for successful transmission from node
i to node j ), (5.3) and (5.4), we have
# m
$1
pij
dij RT (pij
m
)= RTmax ,
P
% m & 1
pkh
dj k m
RI (pkh ) = RImax (k Ijm , k = i, h Tkm ).
P
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Based on the above two constraints, we have the following requirements for
the transmission link i j and the interfering link k h:
) ! d "
m if xijm = 1,
ij
R max P, P
pij T
=0 if xijm = 0.
) d
kj
RImax
P if xijm = 1,
pkh
m
(k Ijm , k = i, h Tkm ).
P if xijm = 0.
Mathematically, these requirements can be rewritten as follows:
% &
dij
(C 1 ) pij
m
P xij , P xij ,
m m
RTmax
% &
dkj
(C 2 ) pkh P 1
m
P xijm (k Ijm , k = i, h Tkm ).
RImax
In addition, for a successful scheduling in the frequency domain, the follow-
ing two constraints must also hold:
Lemma 5.1
If transmission powers on every transmission link and interference link sat-
isfy (C-1 ) and (C-2 ) in the network, then (C-4) and (C-5) are also satisfied.
Proof. We first prove that (C-1 ) and (C-2 ) lead to (C-4). To do this, we let
k = j in (C-2 ). Then (C-2 ) degenerates to pjmh P P xijm since djj = 0.
Suppose that node j is receiving from node i on band m, i.e., xijm = 1. Then,
dj h
pjmh P P xijm = 0. Since pjmh R max P xjmh from (C-1 ), we have that
T
xjmh must be 0. That is, if node j is receiving from node i on band m, then
node j cannot transmit to node h in the same band.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Combining the above two results, we have shown that (C-4) holds.
We now prove that (C-1 ) and (C-2 ) also imply (C-5). Again, the proof is
based on contradiction. Suppose that (C-5) does not hold. Then node j can
receive from two different nodes i and k on the same band m, i.e., xijm = 1 and
m = 1. Note that link k j can be viewed as an interfering link with respect
xkj
to link i j . This corresponds
to letting
h = j in (C-2 ). Then from (C-2 ),
dkj
since xijm = 1, we have pkj
m
RImax
P . Now consider node k is transmitting
dkj
to node j . Then by (C-1 ), we have pkj
m
RTmax P . However, the above two
inequalities cannot hold at the same time since we have RImax > RTmax . This
gives us a contradiction. Thus, the initial assumption that (C-5) does not hold
is incorrect.
The significance of Lemma 5.1 is that, since (C-4) and (C-5) are embedded
in (C-1 ) and (C-2 ), it is sufficient to consider only (C-1 ), (C-2 ), and (C-3) in
the problem formulation. This helps reduce the number of constraints.
=s(l)
j k=
d(l)
fij (l) = fki (l). (5.6)
j Ti kTi
It can be easily verified that once (5.5) and (5.6) are satisfied, (5.7) must also
be satisfied. As a result, it is sufficient to just include (5.5) and (5.6) in the
formulation.
m as link capacity of link i j under p m . In addition to the above
Denote cij ij
flow balance equations at each node i N for session l L, the aggregated
flow rates on each radio link cannot exceed this links capacity. We have
s(l)=
j,d(l)=i % &
gij m
fij (l) m
cij = W log2 1+ p , (5.8)
W ij
lL mMij mMij
% &
dij
qijm Qx m
ij , Qx m
ij (5.10)
R max
% &T
dkj
m
qkh Q 1 Qxijm (k Ijm , k = i, h Tkm ) (5.11)
RImax
s(l)=
j,d(l)=i % &
gij P m
fij (l) W log2 1 + q .
W Q ij
lL mMij
We can re-formulate (5.11) as follows. Note that by (C-3), there is at most one
m = 1. As a result, based on (5.11), there is at most one
h Tkm such that xkh
m m . Thus, (5.11) can be rewritten as
qkh > 0 in hT m qkh
k
% &
dkj
m
qkh Q 1 max Qxijm (k Ijm , k = i).
m R I
hTk
This reformulation will help reduce the number of constraints associated with
(5.11).
Mathematical formulation Putting together the objective function and all
the constraints for power control, scheduling, and flow routing, we have the
following problem formulation:
# $2
qijm
Minimize
Q
iN mMi j Tim
subject to xijm 1 (i N , m Mi ) (5.12)
j Ti m
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
% &
dij
qijm Qxijm 0 (i N , m Mi , j Tim ) (5.13)
RTmax
qijm Qxijm 0 (i N , m Mi , j Tim ) (5.14)
% &
dkj
m
qkh + 1 Qxijm Q
m RImax
hTk
(i N , m Mi , j Tim , k Ijm , k = i)
s(l)=
j,d(l)=i % & (5.15)
gij P m
fij (l) W log2 1+ q 0 (i N , j Ti )
W Q ij
lL mMij
(5.16)
fij (l) = r(l) (l L, i = s(l)) (5.17)
=s(l)
j j Tik=d(l)
fij (l) fki (l) = 0 (l L, i N , i = s(l), d(l)) (5.18)
j Ti kTi
um (q2)2/
ij
(q1)2/
(q0)2/
q0 q1 q2 qk1 qk
qm
ij
% &
dkj
m
qkh + 1 max Qxijm Q (i N , m Mi , j Tim , k Ijm , k = i)
RI
hTkm
s(l)=
j,d(l)=i
fij (l) W vijm 0 (i N , j Ti )
lL mMij
ij , vij 0
um (i N , m Mi , j Tim )
m
where z is the set of all possible values of (x, q) in Problem z. For example,
1 for the original problem (Problem 1) is {(x, q) : 0 xijm 1, 0 qijm
Q}. The above linear relaxation formulation is an LP problem, which can be
easily solved and its objective value serves as a lower bound for Problem z.
is done in Step 9 of Algorithm 5.2, where if a new band m is used for link
i j , then it is necessary to pose a limit on each neighboring transmitter k
m ) -value for
such that the interference from k is negligible. As a result, the (qkh U
k Ijm , k = i, h Tkm may be decreased to satisfy (5.15). Suppose that each
updated (qkh m ) is no less than (q m ) . Then we have a new solution and we
U kh L
can calculate the objective value of this solution. If we can satisfy (5.16) for
all links, then we have a feasible solution. Otherwise (i.e., if (5.16) on any link
cannot be satisfied), we declare that we cannot find a feasible solution and thus
we set the objective value to . The details of this local search algorithm are
presented in Algorithm 5.2.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
ij = (qij )
the relaxation error of a nonlinear discrete term um m 2/ is
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
|um
ij (qij )
m 2/
|;
gij P m
and the relaxation error of a nonlinear discrete term vijm = log2 1 + W Q qij is
% &
m
v log 1 + gij P q m .
ij 2
W Q ij
Among these three types of relaxation errors, we identify the largest one and
choose the corresponding qijm as the partitioning variable. Assuming the value-
set of qijm in Problem z is {q0 , q1 , . . . , qK }, its value-set in Problems z1 and
z2 will be {q0 , q1 , . . . , qijm } and {qijm + 1, qijm + 2, . . . , qK }, respectively.
The new value-set of qijm may narrow other variables value-sets due to con-
straint (5.15). That is, if we increase the lower bound for one q-variable in the
summation in constraint (5.15), the upper bounds of other q-variables in this
summation may be decreased. If these updates make the new value-set of a
variable empty, then the corresponding new subproblem is clearly infeasible.
Again, we only keep feasible subproblems in the problem list.
A feasible subproblem has at least one solution that satisfies constraints
(5.12)(5.15). We can further verify that if we set each qijm = (qijm )L and each
xijm = (xijm )L , then constraints (5.12)(5.15) are satisfied. In particular, con-
straints (5.12)(5.14) hold due to updates in the partitioning process. Since
constraint (5.15) holds by a solution with some qijm values no less than (qijm )L
and some xijm values no less than (xijm )L , decreasing these values will not vio-
late (5.15).
3 7
30
11 15
14
20 2
6
17
10
9
10
1 5
18
13
0
0 10 20 30 40 50
Table 5.2 Each nodes location and available frequency bands for the 20-
node network.
RTmax
have PI = RImax PT W = 50
16 and the maximum transmission power P =
(RTmax ) PT W = 8 106 . We set
= 0.05, which guarantees that the obtained
solution is within 5% of optimality.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 5.3 Source node, destination node, and rate requirement of the five
sessions.
1 7 16 28
2 8 5 12
3 15 13 56
4 2 18 75
5 9 11 29
10
8
Objective value
0
1 3 5 7 9 11 13 15
Q
5.6.2 Results
In this set of results, we apply the proposed solution procedure to the 20-node
network described above for different levels of power control granularity (Q).
Note that Q = 1 corresponds to the case where there is no power control, i.e.,
a node always uses its peak power P for transmission. When Q is sufficiently
large, power control approaches a continuum. Fig. 5.5 shows the results. First,
we note that the granularity of power control has a significant impact on the
optimal objective value. Comparing the case when there is no power control
(Q = 1) and the case of Q = 15, we find that there is nearly a 40% reduction
in the optimal objective value. Second, although the optimal objective value
is a nonincreasing function of Q, when Q becomes sufficiently large (e.g., 10
in this network setting), further increase in Q does not have much reduction
in the optimal objective value. This suggests that, for practical purposes, the
number of required power control levels does not need to be a large number.
The rest of our results in this section are for Q = 10. For transmission
power, we obtain the following results:
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
40
12
3 7
30
11 15
14
20
2 6 17
10
9
10
5
1
18 13
0
0 10 20 30 40 50
1
q9,11 = 3, 2
q12,16 = 4, 3
q7,12 = 3,
4 = 4,
q2,1 q2,1 = 4,
5 q5,18 = 2,
6
7
q17,13 = 4, 8 = 5, q 8
q8,3 14,17 = 1,
9
q1,5 = 1, q15,14
9
= 1,
q3,2 = 5.
10
another observation (routing), we see that, for Session 2 (node 8 to node 5), the
routing path is 8 3 2 1 5. Here, a shorter path 8 3 6 5 is
not used. This is because path 8 3 6 5 is interfered by transmissions
on other paths. The optimal solution tends to choose paths that are not close
to each other. Finally, for Session 3 (node 15 to node 13), the routing path is
15 14 17 13, while a shorter path 15 17 13 is not used. This
is because node 15 can use a smaller transmission power to transmit to a closer
neighboring node 14 with a smaller interference footprint. This allows link
1 5 to be active on the same band 9 with link 15 14.
This is the first chapter of the second part of this book, which presents some
methods to develop (1 )-optimal solutions. In this chapter, we presented
the branch-and-bound framework and showed how it could be applied to solve
discrete and combinatorial optimization problems. Such problems are typi-
cally considered most difficult in nonconvex optimization, and the branch-and-
bound framework offers a general purpose and effective approach. The effec-
tiveness of branch-and-bound resides in the careful design of each component
in its framework, such as computation of a lower bound, local search of an
upper bound, and selection of partitioning variables (in the case of a minimiza-
tion problem). It should be noted that the worst-case complexity of a branch-
and-bound-based method remains exponential, although a judicious design of
each component could achieve reasonable computational times in practice.
In the case study, we considered a per-node power control problem for a
multi-hop CRN. This problem has a large design space that involves a tight
coupling relationship among power control, scheduling, and flow routing,
which is typical for a cross-layer optimization problem. We developed a math-
ematical model and a problem formulation, which is a mixed-integer nonlinear
programming (MINLP) problem. We showed how to apply the branch-and-
bound framework to design a solution procedure. Under the branch-and-bound
framework, we showed how to derive a linear relaxation to compute a lower
bound, how to perform a local search to find a feasible solution (upper bound),
and how to select suitable partitioning variables. Despite its worst-case expo-
nential complexity, the solution procedure that we developed here is a viable
approach to solve the per-node power control problem for a multi-hop CRN
(as demonstrated in the numerical examples).
5.8 Problems
5.2 In the case study, we employed the so-called protocol model. Read
[152] and discuss the pros and cons of this model.
5.3 Show that for a session, if flow balance holds at the source node
[Eq. (5.5)] and at the relay nodes [Eq. (5.6)], then flow balance also holds
at this sessions destination node [Eq. (5.7)].
5.4 Referring to the RHS of (5.8), explain why there is no interference term
in the denominator inside the log function to calculate SINR.
5.5 In this chapter, we used the so-called network-wide bandwidth-footprint-
product (BFP) as the objective function. However, it is possible that two or
more neighboring nodes transmitting on the same frequency band may have
a partial overlap of their footprints. As a result, the overlapped area will be
counted multiple times in the objective function. Give a justification on why
this still makes sense.
5.6 In the formulation of the original problem, what constraints are associ-
ated with a single layer? What constraints couple multiple layers together?
5.7 In the formulation of the original problem, we discretize transmission
power from a continuous variable to a discrete variable. Then in the relax-
ation step of the branch-and-bound solution procedure, we change the discrete
transmission power variable to a continuous variable. Explain the purpose of
this back and forth change between continuous and discrete variables for the
transmission power.
5.8 We showed how to construct a linear relaxation for the discrete term
(qijm )2/ . Provide the details for deriving a linear relaxation for the discrete
gij P m
term log2 1 + W Q ij in the problem formulation.
q
5.9 There is another approach to obtain a linear relaxation for the discrete
term (qijm )2/ , where qijm {q0 , q1 , . . . , qK }. Introduce binary variables yijmk ,
where yijmk = 1 if qijm = qk and yijmk = 0 otherwise. Thus, K k=0 yij = 1 and
mk
qijm = K mk m
k=0 yij qk . We also introduce a new variable uij for (qij )
m 2/ and have
K
umij =
mk 2/ . A linear relaxation can be obtained by relaxing y mk
k=0 yij (qk ) ij
to a continuous variable in [0, 1]. Compare this approach and the linear relaxa-
tion approach in Section 5.5.1 and analyze which provides a tighter relaxation.
5.10 In local search, why do we initialize qijm at (qijm )L and increase it
upward? Comment on the strategy of starting qijm at (qijm )U and decreasing
it downward.
5.11 In local search, we focus on increasing the power values (see Steps 5,
8, and 11 in Algorithm. 5.2) to ensure that constraint (5.16) holds for each link
i j . Explain how other constraints in the original problem hold.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
5.12 In the selection of partitioning variables, why are the x-variables more
important than the q-variables? After branching a problem into two new
subproblems, how can we revise the value-sets of the other x- and q-variables
by using constraints (5.12)(5.15)?
5.13 When we examine the relaxation errors for the q-variables, why can
ij (qij )
|um m 2/ | be greater than zero?
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CHAPTER
6 Reformulation-linearization
technique and applications
new variable X{1,2} to represent x1 x2 and add suitable linear constraints for
this new variable. Suppose that we have derived (or are given) the lower
and upper bounds for variables x1 and x2 , i.e., (x1 )L x1 (x1 )U and
(x2 )L x2 (x2 )U , respectively. Then the following so-called bound-factor
constraints must hold:
[x1 (x1 )L ] [x2 (x2 )L ] 0,
[x1 (x1 )L ] [(x2 )U x2 ] 0,
[(x1 )U x1 ] [x2 (x2 )L ] 0,
[(x1 )U x1 ] [(x2 )U x2 ] 0.
Substituting X{1,2} for x1 x2 in the above constraints, we obtain
(x1 )L x2 + (x2 )L x1 X{1,2} (x1 )L (x2 )L , (6.1)
(x1 )U x2 + (x2 )L x1 X{1,2} (x1 )U (x2 )L , (6.2)
(x1 )L x2 + (x2 )U x1 X{1,2} (x1 )L (x2 )U , (6.3)
(x1 )U x2 + (x2 )U x1 X{1,2} (x1 )U (x2 )U . (6.4)
Therefore, a linear relaxation for the original problem can be obtained by
replacing the bilinear term x1 x2 by X{1,2} throughout the problem and adding
the above four linear constraints. Note that (6.1) (6.4) are valid restrictions
that relate the original variables x1 and x2 and the new RLT-variable X{1,2} .
In particular, X{1,2} = x1 x2 holds true when either x1 or x2 equals its corre-
sponding lower or upper bound value. This feature, which holds in a more
general form (see [141]), plays a decisive role in the tightness of RLT.
Next, we show how to obtain a linear relaxation for x12 in a nonconvex pro-
gram. Viewing x12 as x1 x1 , we derive a linear relaxation using RLT, by intro-
ducing a new variable X{1,1} to represent x12 , where the subscript of X repeats
the index 1 twice corresponding to the exponent in x12 , and by adding suitable
linear constraints for relating the new variable X{1,1} to the original variable
x1 . This is done similar to the above, by regarding x12 as a special case of x1 x2
when x2 x1 . Therefore, a linear relaxation for the original problem can be
obtained by replacing all occurrences of x12 within the polynomial program by
X{1,1} and by adding the following three bound-factor linear constraints. Note
that for this special case, both (6.2) and (6.3) become (6).6:
2(x1 )L x1 X{1,1} [(x1 )L ]2 , (6.5)
[(x1 )U + (x1 )L ] x1 X{1,1} (x1 )U (x1 )L , (6.6)
2(x1 )U x1 X{1,1} [(x1 )U ]2 . (6.7)
This strategy can be used in a likewise fashion for constructing a linear relax-
ation via RLT for a general monomial ni=1 (xi )ci in variables xi , where the
ci -exponents are integer constants for 1 i n. Although the general RLT
process is described in [141], we present a sequential quadrification process
[143] here for the sake of simplicity. This strategy adopts a two-step approach
for the linearization process. In the first step, we introduce new variables and
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
identities to represent ni=1 (xi )ci using quadratic terms. In the second step, we
introduce RLT-variables as above to represent the quadratic terms and generate
the related RLT bound-factor constraints in order to derive a linear relaxation.
The following example illustrates the general idea:
Example 6.1
Suppose that we have a nonlinear term x1 x25 x3 in a nonconvex program
with variables x1 , x2 , and x3 . We begin by factoring x1 x25 x3 into quadratic
relationships as follows by defining new variables (note that this is not a
unique representation [143]):
X{2,2} = x22 , (6.8)
X{2,2,2,2} = 2
X{2,2} , (6.9)
X{1,2} = x1 x2 , (6.10)
X{1,2,2,2,2,2} = X{1,2} X{2,2,2,2} , (6.11)
X{1,2,2,2,2,2,3} = X{1,2,2,2,2,2} x3 . (6.12)
Note that in terms of the original variables x1 , x2 , and x3 , we have that
X{1,2,2,2,2,2,3} = x1 x25 x3 . Hence, we replace x1 x25 x3 in the given polyno-
mial program with the new variable X{1,2,2,2,2,2,3} . Furthermore, as before,
we include suitable linearized bound-factor constraints to represent the
quadratic relationships in the above equalities. For example, for (6.8), we
include the following three linear constraints, as in (6).5 (6).7 above:
2(x2 )L x2 X{2,2} [(x2 )L ]2 ,
[(x2 )U + (x2 )L ] x2 X{2,2} (x2 )U (x2 )L ,
(x2 )U x2 X{2,2} [(x2 )U ]2 .
The similar three linear constraints for representing (6.9) are left as a home-
work exercise (note that implied lower and upper bounds would need to be
derived for the variables X{1,2} , X{2,2,2,2} , and X{1,2,2,2,2,2} ). For (6.10), we
generate the following four bound-factor linear constraints as in (6.1)
(6.4):
(x1 )L x2 + (x2 )L x1 X{1,2} (x1 )L (x2 )L ,
(x1 )U x2 + (x2 )L x1 X{1,2} (x1 )U (x2 )L ,
(x1 )L x2 + (x2 )U x1 X{1,2} (x1 )L (x2 )U ,
(x1 )U x2 + (x2 )U x1 X{1,2} (x1 )U (x2 )U .
The similar four linear constraints for representing each of the relation-
ships (6.11) and (6.12) are relegated to the exercises. This produces a linear
relaxation for the original problem with five additional variables and 24
additional linear constraints.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
For a more comprehensive discussion and surveys of RLT and related theo-
retical results, we refer the readers to [146; 148; 149]. In the rest of this chapter,
we will show how RLT can be applied to solve problems in a wireless network.
6.3.1 Power control, scheduling, and their relationship in the SINR model
Power control on each transmitting node at the physical layer affects SINR at
a receiving node. These SINR values in turn will affect scheduling decisions at
the link layer. That is, if a node is scheduled to receive, then its SINR must be
at least Smin (minimum threshold requirement). Therefore, power control and
scheduling are tightly coupled via SINR and cannot be modeled separately.
Scheduling at a node can be done either in the frequency domain or time
domain. In this chapter, we consider scheduling in the frequency domain in the
form of assigning frequency bands (channels). Note that a time domain-based
formulation can be done in a similar fashion.
In the SINR model, there may still be concurrent transmissions within
the same channel (and thus interference). Denote scheduling variables xijm as
follows:
'
1 if node i transmits data to node j on band m,
xijm =
0 otherwise.
We assume that a node can use a band for transmission (or reception) to (or
from) only one other node. That is,
j :mMj
k
k:mM
m
xki + xijm 1 (i N , m Mi ). (6.13)
kN ,k=i j N ,j =i
For power control, we assume that the transmission power at a node can
only be tuned to a finite number of levels between 0 and Pmax . To model this
discrete power control, we introduce an integer parameter Q that represents
the total number of power levels to which a transmitter can be adjusted,
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Symbol Definition
m
cij Capacity of link i j under pij m
ij qm
gij Q Pmax
sijm = k:mMk h:mMh m
qkh
(i, j N , i = j, m Mij ), (6.15)
W + kN ,k=i hN ,h=k gkj Q Pmax
where is the ambient Gaussian noise density and gij is the propagation loss
from node i to node j .
Note that, in theory, for any small SINR, the corresponding capacity is still
positive (by Shannons capacity formula). But in practice, if SINR is too small,
then the achieved capacity will also be very small. In this case, such a weak
link will not be very useful to carry traffic flow. Thus, we may use a threshold
to remove such weak links from consideration. In this regard, we introduce
a threshold for SINR, i.e., a transmission from node i to node j on band m
is considered successful if and only if sijm Smin . We thus have the following
coupling relationship for scheduling (xijm ) and SINR (sijm ):
k:M ki =
fki (l) = K r(l) (l L, i = d(l)). (6.19)
kN ,k=i
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
In addition to the above flow balance equations at each node i N for ses-
sion l L, we impose a constraint to assure that the aggregated flow rates on
each radio link do not exceed this links capacity, i.e., for a link i j , we
have
s(l)=
j,d(l)=i
fij (l) W log2 (1 + sijm ) (i, j N , i = j, Mij = ).
lL mMij
(6.20)
This constraint shows the coupling relationship between flow routing and
SINR.
Maximize K
subject to Constraints (6.13)(6.20)
xijm {0, 1}, qijm {0, 1, 2, . . . , Q}, tim , sijm 0 (i, j N , i = j, m Mij )
6.4 Reformulation
A formulation like the one in Section 6.3.3 is the first step in formulating our
cross-layer optimization problem. But it is in a rather raw form and more
work needs to be done to reformulate it into a more compact form that is
amenable to solution development. In this section, we analyze each constraint
in detail and perform some necessary and important reformulations.
qijm
Q Pmax
gij
sijm = k:mMk h:mMh qm
W + kN ,k=i hN ,h=k gkj Qkh Pmax
gij qijm
= W Q k:mMk h:mMh
(i, j N , i = j, m Mij ).
Pmax +
m
kN ,k=i hN ,h=k gkj qkh
This is equivalent to
W Q m k h:mM
k:mM h
s + m m
gkj qkh sij gij qijm = 0
Pmax ij
kN ,k=i hN ,h=k
h
h:mM
tkm = m
qkh (k N , m Mk ). (6.23)
hN ,h=k
W Q m k
k:mM
sij + gkj tkm sijm gij qijm = 0 (i, j N , i = j, m Mij ), (6.24)
Pmax
kN ,k=i
which now only involves a single sum of nonlinear terms tkm sijm .
Similar to (6.14), the constraint described in (6.16) is not suitable for math-
ematical programming. It can be shown that (6.16) can be eliminated if we
have (6.21), (6).24, and the following new constraint:
Maximize K
k:mMk m j :mMj m
subject to kN ,k=i xki + j N ,j =i xij 1 (i N , m Mi )
m xm 0
qij (i, j N , i = j, m Mij )
ij
m Qx m 0
qij (i, j N , i = j, m Mij )
ij
j :mMj m m
j N ,j =i qij ti = 0 (i N , m Mi )
W Q m k:mMk m m m
Pmax sij + kN ,k=i gkj tk sij gij qij = 0 (i, j N , i = j, m Mij )
m sm 0
Smin xij (i, j N , i = j, m Mij )
ij
j :Mij =
j N ,j =i fij (l) r(l)K = 0 (l L, i = s(l))
j :Mij = k:Mki =
j N ,j =i,s(l) fij (l) kN ,k=i,d(l) fki (l) = 0 (l L, i N , i = s(l), d(l))
s(l)=j,d(l)=i m
lL fij (l) mMij W log2 (1 + sij ) 0 (i, j N , i = j, Mij = )
For the optimization problem in Table. 6.2, K, xijm , qijm , tim , sijm , and fij (l) are
optimization variables and Q, , W, Smin , Pmax , gij , and r(l) are constants.
This formulation is a mixed-integer nonlinear programming (MINLP), which
is NP-hard in general [46]. In Section 6.5.1, we first analyze the intricate rela-
tionship among the variables and identify the core variables among all the vari-
ables. We show that the dependent variables can be derived once these core
variables are fixed. We call the optimization space for the core variables the
core optimization space. In Section 6.5.2, we present the main algorithm on
how to determine an optimal solution in the core optimization space. Several
key components in the main algorithm are described in Sections 6.5.3, 6.5.4,
and 6.5.5.
inter-dependent. In particular, we find that the xijm - and qijm -variables are core
variables and the other variables {tim , sijm , fij (l), K} can all be derived based on
these core variables. How to derive these dependent variables based on the core
variables is left as a homework problem. As a result, we can focus our study
on an optimization space defined by the core variables, xijm and qijm , which is a
much smaller space.
6.5.2 A solution
In this section, we describe a solution procedure based on the branch-and-
bound framework. Recall that under branch-and-bound, we aim to provide a
(1 )-optimal solution, where is a small positive constant that reflects our
desired minimal accuracy (1 ) in computing the final solution. A branch-
and-bound framework for a minimization problem was presented in Sec-
tion 5.1. In this chapter, we need to solve a maximization problem, which
follows a similar scheme and is described in Algorithm 6.1.
Since our core optimization space is finite (with a finite number of core
variables xijm and qijm , where each core variable has a finite integer value-set),
the branch-and-bound algorithm is guaranteed to converge in a finite number
of iterations. Several components (i.e., determining upper and lower bounds,
and partitioning subproblems) in the main algorithm are yet to be developed.
These components should exploit problem-specific structures to optimize per-
formance. In the rest of this section, we show how these components can be
designed.
For the nonlinear term log2 (1 + sijm ) = ln12 ln(1 + sijm ), we propose to
employ three tangential supports for ln(1 + sijm ), which yields a relaxation
for the underlying convex hull linear representation (see Fig. 6.1). Sup-
pose that sijm is bounded by (sijm )L sijm (sijm )U . We introduce a variable
m = ln(1 + s m ), and bound the convex hull of the region defined by the
cij ij
curve cijm over (s m ) s m (s m )
ij L ij ij U by using four segments as displayed
in Fig. 6.1, where the segments I, II, and III are tangential supports and
the segment IV is the chord. In particular, the three segments I, II, and III
are tangential at points (1 + (sijm )L , ln(1 + (sijm )L )), (1 + , ln(1 + )), and
(1 + (sijm )U , ln(1 + (sijm )U )), where
II
cm=In(1+sm)
IV ij ij
sm +1
ij
(sm ) +1 b+1 (sm ) +1
ij L ij U
is the s-value corresponding to the intersection point of the segments I and III
(see Fig. 6.1); and segment IV is the chord that joins points (1 + (sijm )L , ln(1 +
(sijm )L )) and (1 + (sijm )U , ln(1 + (sijm )U )). The convex region defined by the
four segments can be described by the following four linear constraints:
[1 + (sijm )L ] cij
m s m [1 + (s m ) ] [ln(1 + (s m ) ) 1] + 1,
ij ij L ij L
m s m (1 + ) [ln(1 + ) 1] + 1,
(1 + ) cij ij
[1 + (sijm )U ] cij
m s m [1 + (s m ) ] [ln(1 + (s m ) ) 1] + 1,
ij ij U ij U
Maximize K
k:mMk m j :mMj m
subject to kN ,k=i xki + j N ,j =i xij 1 (i N , m Mi )
m xm 0
qij (i, j N , i = j, m Mij )
ij
m Qx m 0
qij (i, j N , i = j, m Mij )
ij
j :mMj m m
j N ,j =i qij ti = 0 (i N , m Mi )
W Q m k:mMk m m
Pmax sij + kN ,k=i gkj uij k gij qij = 0 (i, j N , i = j, m Mij )
m sm 0
Smin xij (i, j N , i = j, m Mij )
ij
j :Mij =
j N ,j =i fij (l) r(l)K = 0 (l L, i = s(l))
j :Mij = k:Mki =
j N ,j =i,s(l) fij (l) kN ,k=i,d(l) fki (l) = 0 (l L, i N , i = s(l), d(l))
s(l)=j,d(l)=i W m
lL fij (l) mMij ln 2 cij 0 (i, j N , i = j, Mij = )
m
Linear constraints for cij (i, j N , i = j, m Mij )
tim , sij
m , cm , um 0
ij ij k (i, j, k N , i = j, m Mij , m Mk )
(x, q) z
to Problem z can serve to provide a lower bound, it is desirable to find one that
offers a tight lower bound, i.e., with an objective value close to the optimal
value. Such a feasible solution (denoted by z ) can be found by searching the
neighborhood of LP (z), a process that is called local search.
A local search algorithm begins with an initial feasible solution. Such a solu-
tion may be far away from the optimum and may not provide a tight lower
bound. But we can iteratively improve the current solution to achieve a bet-
ter lower bound until we can no longer improve (increase) the lower bound,
whence we terminate this search process.
To obtain an initial feasible solution, we set xijm = (xijm )L for scheduling
and qijm = (qijm )L for power control. Then we can compute SINR value sijm
by (6).15. When an SINR value is larger than or equal to Smin , the achieved
capacity is W log2 (1 + sijm ). Otherwise (i.e., SINR < Smin ), the transmission is
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
considered unsuccessful. Note that although the flow rates fij (l) in the relaxed
solution LP (z) guarantee flow balance at each node, such flow rates may
exceed the capacities on some links under the initial xijm - and qijm -values. To find
feasible flow rates for the current xijm - and qijm -values, we compare the achiev-
able link capacity (under this current solution) to the aggregated flow rates
fij (l) on each link i j by computing the ratio between the two (denoted as
ij ) as follows:
mMij W log2 (1 + sij )
m
ij = s(l)=j,d(l)=i . (6.26)
lL fij (l)
If ij < 1 for some link i j , then the aggregated flow rates exceed the
link capacity and the link capacity constraint on i j is violated. In this
case, we need to scale down the flow rates on link i j (to satisfy the link
capacity constraint) and the flow rates on all other links (to maintain flow
balance in the network) by a value ij . On the other hand, we want to
have a as large as possible so as to maximize the scaling factor (our objec-
tive). Such a value is the bottleneck value ij among all links (denoted as
min = min{ij : i, j N , i = j, Mij = }). We now have a complete solu-
tion min fij (l), (xijm )L , (qijm )L for routing, scheduling, and power control,
respectively. The achieved objective is min K, where K is the objective value
in the relaxed solution LP (z).
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
In the next iteration, we aim to improve the current solution. Note that if
we can increase min , then the current solution is improved. Suppose that link
i j is the link with ij = min . To increase ij , we try to increase the trans-
mission power qijm on some band m under the constraint qijm (qijm )U . Based
on the constraints in Table 6.2, we may update the values of other variables to
maintain feasibility. For example, by the first constraint in Table 6.2, we need
to increase xijm from 0 to 1 if qijm is increased to a positive value. As a con-
sequence of increasing qijm , the interference with other transmissions on band
m is increased and thus the achieved capacities on other links are decreased.
Thus, qijm can be successfully increased only if for any other link k h, its
updated kh does not fall below the current min . If the current solution can
be improved (with a larger min ), then we continue to the next iteration and
try for further improvement. Otherwise, the local search algorithm terminates.
The pseudocode for this local search algorithm is given in Algorithm 6.2.
the value-set for this variable is set to either 0 or 1 in the two respective
subproblems.
We further observe that by fixing the value of xijm , some other variables may
also be fixed. Note that by (6.13), a node can only receive from or transmit to
one node on the same band. Based on this observation, if the value of xijm is
set to 1, then we have xki m = 0 for k N , k = i, if m M ; x m = 0 for p
k ip
N , p = j , if m Mp ; xjmh = 0 for h N , h = j , if m Mh ; and xqj m =0
In this section, we present some numerical results for the solution procedure.
Table 6.4 Location and available frequency bands at each node for
a 20-node network.
1 (0.1, 9.9) 1, 2, 3, 4, 7, 8, 9, 10
2 (29.2, 31.7) 1, 2, 3, 4, 5, 7, 8, 10
3 (3, 31.1) 1, 4, 5, 6
4 (11.8, 40.1) 1, 2, 3, 4, 6, 9, 10
5 (15.8, 9.7) 1, 2, 3, 5, 6, 8, 9
6 (16.3, 19.5) 3, 5, 6, 8, 9
7 (0.6, 27.4) 1, 4, 8, 9, 10
8 (22.6, 40.9) 1, 2, 3, 5, 7, 9, 10
9 (35.3, 10.3) 2, 9
10 (31.9, 19.6) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
11 (28.1, 25.6) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
12 (32.3, 38) 1, 8, 9, 10
13 (47.2, 2.6) 3, 5, 10
14 (44.7, 15) 2, 3, 6, 7, 8
15 (44.7, 24) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
16 (47.9, 43.8) 1, 3
17 (46.4, 16.8) 1, 7, 9
18 (11.5, 12.2) 2, 5, 6, 10
19 (28.2, 14.8) 4, 5, 6, 7, 8, 9, 10
20 (2.5, 14.5) 1, 7, 10
For our proposed algorithm, we set 0.1, which guarantees that the solu-
tion is at least 90% optimal.
We note that the brute-force approach cannot solve the problem even for
the 20-node CRN. The solution space of the capacity problem in Section 6.3.3
includes all possible sets of values for (xijm , qijm , K, fij (l)). Thus, the number
of solutions examined in the brute-force approach is clearly more than the
number of all possible sets of values for the qijm -variables. For the 20-node
CRN, the number of qijm -variables is about 20 (20 1) 5 = 1900, where 5
is an approximation of the average number of available bands on a link. Each
qijm -variable has (Q + 1) or 11 possible values. Thus, the number of all possi-
ble sets of values for the qijm -variables is about 111900 . Therefore, for the 20-
node CRN, the number of solutions examined in the brute-force approach is
at least 111900 . Even if each solution can be examined in 106 second, the run
time is 111900 106 > 111894 seconds, which is 111894 /(365 24 60 60) >
111894 /108 > 111886 years! Therefore, a brute-force approach cannot be used
to solve our problem except for toy-sized instances.
6.6.2 Results
For the 20-node network with five sessions, our solution achieves a scal-
ing factor K = 13.24. Based on the minimum rate requirement r(l) in
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 6.5 Source node, destination node, and minimum rate requirement
for each session in the 20-node network.
Session l Source node s(l) Dest. node d(l) Min. rate req. r(l)
1 16 10 9
2 18 3 1
3 12 11 4
4 13 17 3
5 15 6 2
35
3 2
30
7
25 11
15
6
20 10
17
20
15 18 14
19
10 1 5 9
5
13
0
0 5 10 15 20 25 30 35 40 45 50
Table 6.5, the corresponding flow rates K r(l) for the five sessions are
119.16, 13.24, 52.96, 39.72, 26.48, respectively. Fig. 6.2 shows the routing
topology for the final solution. The flow rates for each session on the links
along its path are as follows:
Session l = 1: f16,12 (1) = 119.16, f12,8 (1) = 103.30, f12,11 (1) = 15.86,
f8,2 (1) = 103.30, f11,10 (1) = 15.86, f2,10 (1) = 103.30;
Session l = 2: f18,20 (2) = 13.24, f20,1 (2) = 13.24, f1,7 (2) = 13.24,
f7,3 (2) = 13.24;
Session l = 3: f12,11 (3) = 52.96;
Session l = 4: f13,14 (4) = 39.72, f14,17 (4) = 39.72;
Session l = 5: f15,19 (5) = 26.48, f19,6 (5) = 26.48.
It is easy to verify that flow balance holds at all nodes. Note that flow splitting
and multi-path routing are used for Session 1, which has the largest rate
requirement.
Our solution also prescribes the scheduling variables xijm as follows, where
we indicate only the nonzero xijm -variables:
Band m = 1: x7,3 1 = 1, x 1
16,12 = 1;
Band m = 2: x8,2 2 = 1;
Band m = 3: x13,14
3 = 1;Band m = 4: x1,7
4 = 1, x 4
2,10 = 1;
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Band m = 5: x11,10
5 = 1;
Band m = 6: x15,19 = 1;
6
Band m = 7: x14,17
7 = 1, x20,1
7 = 1;
Band m = 8: x12,11 = 1;
8
Band m = 9: x12,8
9
= 1, x19,6
9
= 1;
Band m = 10: x18,20 = 1.
10
Band m = 3: q13,14
3 = 2;
Band m = 4: q1,7
4 = 7, q 4
2,10 = 2;
Band m = 5: q11,10 = 1;
5
Band m = 6: q15,19
6 = 9;
Band m = 7: q14,17
7 = 1, q20,1
7 = 1;
Band m = 8: q12,11 = 3;
8
Band m = 9: q12,8
9
= 1, q19,6
9
= 3;
Band m = 10: q18,20 = 1.
10
Note that the same frequency band may be used by concurrent transmissions.
1 = 1 and x 1
For example, since x7,3 16,12 = 1, we have that both nodes 7 and
16 are transmitting on band 1. Such concurrent transmissions are allowed as
long as the SINR at each receiving node is no less than Smin . For example,
g q 1
the SINR at the receiving node 12 on band 1 is s16,121 = W Q16,12 16,12 1 =
Pmax +g7,12 q7,3
(1.303105 )7
= 4.22, which is larger than Smin = 3. Thus, the trans-
2.083105 +(8.011107 )1
mission 16 12 on band 1 is successful. Following the same token, we can
compute all SINR sijm -values as follows:
Band m = 1: s7,31 = 118.47, s 1
16,12 = 4.22;
Band m = 2: s8,2 = 5.84;
2
Band m = 3: s13,14
3 = 3.75;
Band m = 4: s1,74 = 3.33, s 4
2,10 = 3.14;
Band m = 5: s11,10 = 18.87;
5
Band m = 6: s15,19
6 = 3.39;
Band m = 7: s14,17 = 1261.14, s20,1
7 7 = 65.46;
Band m = 8: s12,11 = 4.90;
8
Band m = 9: s12,8
9
= 3.56, q19,6
9
= 4.74;
Band m = 10: s18,20 = 6.45.
10
As expected, we see that the achieved SINR at each receiving node on each
band is larger than Smin = 3 in the derived solution.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
For each link, we can further verify that the flow rates on this link do
not exceed its capacity. For example, for link 16 12, there is a flow rate
f16,12 (1) = 119.16 on this link. The achieved capacity is W log2 (1 + s16,12
1 )=
W log2 (1 + 4.22) = 119.21.
The above results are for the 20-node network. The results for the 30-
and 50-node networks are similar and we abbreviate our discussion. For
the 30-node five-session network, the available bands at each node and the
location of each node are displayed in Table 6.6. The source, destination,
and minimum rate requirement for each session are shown in Table 6.7.
The routing topology for the derived solution is depicted in Fig. 6.3. The
achieved scaling factor is 31.18. Accordingly, based on the minimum rate
requirement r(l) in Table 6.7, the flow rates K r(l) for the five sessions
Table 6.6 Location and available frequency bands at each node for a
30-node network.
Table 6.7 Source node, destination node, and minimum rate requirement
for each session in the 30-node network.
Session l Source node s(l) Dest. node d(l) Min. rate req. r(l)
1 16 28 4
2 24 11 7
3 13 1 1
4 19 29 8
5 26 15 1
Table 6.8 Flow rates for each session in the 30-node five-session network.
are 124.72, 218.26, 31.18, 249.44, 31.18, respectively. The flow rates for each
session on links along its path are shown in Table 6.8. The nonzero schedul-
ing variables xijm and the nonzero power control variables qijm are specified in
Tables 6.9 and 6.10, respectively.
For the 50-node ten-session network, the available bands at each node and
the location of each node are displayed in Table 6.11. The source, destination,
and minimum rate requirement for each session are specified in Table 6.12.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 6.10 Transmission power levels for the 30-node five-session network.
The routing topology for the derived solution is depicted in Fig. 6.4. The
achieved scaling factor is 13.36. The detailed solution regarding power control,
scheduling, and routing are similar to the 20- and 30-node networks, and are
hence omitted.
Table 6.11 Location and available frequency bands at each node for a 50-node
network.
1 (11.1, 21.7) 2, 3, 4, 8, 25
2 (0.1, 4) 6, 7, 10, 13, 14, 20, 23, 24, 26, 28
3 (7.2, 16.6) 6, 10, 14, 20, 23, 24, 26
4 (11, 32.2) 6, 7, 10, 13, 14, 20, 23, 24, 26, 28
5 (16.3, 3.6) 10, 13, 14, 20, 23
6 (14.5, 24.7) 8, 11, 25
7 (14.9, 13.7) 5, 9, 12, 16, 17, 18, 22, 27, 29, 30
8 (19.5, 14.9) 7, 24, 28
9 (26.6, 13.4) 1, 19, 21, 25
10 (22.5, 29.3) 1, 3, 4, 8, 11, 15, 19
11 (24.6, 40.5) 3, 8, 25
12 (38.4, 13.1) 2, 8, 11, 15
13 (4, 3.9) 9, 12, 16, 22, 27, 29, 30
14 (6.1, 18.6) 9, 12, 16, 17, 18, 22, 27, 30
15 (38.5, 22.6) 2, 4, 11, 15, 19, 21, 25
16 (1.2, 24.3) 5, 9, 12, 17, 22, 29, 30
17 (4.9, 42.3) 5, 27
18 (18.5, 1.4) 5, 9, 12, 17, 18, 27, 30
19 (16.9, 29.1) 3, 4, 10, 11, 12, 15
20 (33.5, 10.4) 7, 13, 14, 20, 23, 24, 26, 28
21 (25.6, 12.8) 6, 7, 20, 23, 24, 28
22 (45.2, 45.5) 2, 8, 15, 19
23 (43.6, 22.7) 1, 2, 3, 4, 11, 15, 19, 21
24 (10.6, 40.5) 4, 15, 19, 21, 25
25 (18.2, 32.7) 9, 12, 18, 22, 27
26 (25.2, 27.2) 10, 14, 20, 24, 26
27 (22.5, 42.2) 5, 9, 12, 16, 18, 27, 29, 30
28 (30, 31.5) 6, 13, 24, 26, 28
29 (35, 22.1) 6, 10
30 (25.7, 6.2) 5, 9, 12, 17, 18, 22, 27, 29, 30
31 (34.1, 12.4) 9, 12, 16, 17, 30
32 (26.4, 30) 5, 9, 12, 16, 17, 18, 22, 27, 29, 30
33 (14.1, 40.7) 1, 2, 25
34 (34.4, 46.5) 9, 17, 18, 30
35 (19, 22.5) 1, 6, 7, 10, 13, 14, 20, 23, 24, 28
36 (39.9, 25.1) 6, 13, 14, 20, 23, 24, 26, 28
37 (20.3, 18.2) 1, 2, 3, 4, 8, 11, 15, 19, 21, 27
38 (10, 20.5) 6, 7, 10, 13, 14, 20, 23, 24, 26, 28
39 (20.5, 21.4) 1, 2, 3, 4, 8, 11, 15, 19, 21, 25
40 (37.1, 28.6) 7, 10, 13, 14, 20, 23, 24, 26
41 (44.1, 16.1) 1, 15, 21
42 (41.1, 6) 9, 29
43 (43, 18.8) 5, 9, 12, 16, 18, 22
44 (45.4, 24.2) 9, 12, 16, 17, 18, 30
45 (36.2, 41.2) 5, 9, 17, 27, 29, 30
46 (27.5, 32.3) 12, 16, 17, 18, 29, 30
47 (47.8, 13.8) 22, 27, 29, 30
48 (8.9, 14.8) 5, 30
49 (6.8, 6.2) 5, 9, 12, 16, 17, 27, 30
50 (11.7, 35.8) 1, 2, 3, 4, 8, 11, 15, 19, 21, 25
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 6.12 Source node, destination node, and minimum rate requirement
of each session in the 50-node network.
Session l Source node s(l) Dest. node d(l) Min. rate req. r(l)
1 21 4 4
2 5 26 7
3 19 20 6
4 33 6 10
5 37 10 9
6 23 11 2
7 25 46 3
8 42 43 9
9 44 27 8
10 47 30 1
35 s(7) d(7)
d(1)
30 d(5)
s(3)
d(2)
25 d(4) s(9)
s(6)
20
s(5) d(8)
15
s(1) s(10)
10
d(3)
5 s(8)
s(2) d(10)
0
0 5 10 15 20 25 30 35 40 45 50
can be applied to any polynomial term of the form ni=1 (xi )ci in variables xi ,
where the ci -exponents are constant integers. Given such generality, RLT is a
powerful tool in deriving tight linear relaxations.
As a case study, we considered a throughput maximization problem in a
multi-hop CRN under the SINR model. We developed a mathematical formu-
lation for joint optimization of power control, scheduling, and flow routing. We
presented a solution procedure based on the branch-and-bound framework and
applied RLT in deriving tight linear relaxations for a product of variables. In
this case study, we also learned how to identify the core optimization space for
the underlying problem and how to exploit different physical interpretations of
the core variables in developing a solution.
6.8 Problems
6.1 For the first simple example of RLT in Section 6.1, we showed that a
nonlinear (bilinear) term x1 x2 can be represented by a single variable X{1,2}
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
and four linear constraints (6.1) (6.4). Show that X{1,2} = x1 x2 holds true
when either x1 or x2 equals to its corresponding lower or upper bound value.
6.2 For Example 6.1, please show the three linear constraints for represent-
ing the relationship X{2,2,2,2} = X{2,2}
2 , and the four linear constraints for rep-
resenting X{1,2,2,2,2,2} = X{1,2} X{2,2,2,2} and X{1,2,2,2,2,2,3} = X{1,2,2,2,2,2} x3 ,
respectively. Derive appropriate lower and upper bounds on variables as nec-
essary for this representation.
6.3 In the formulation of the multi-hop CRN problem, describe the set of
constraints associated with each layer. Which set of constraints couples multi-
ple layers together?
6.4 By introducing new variables tkm , we can replace (6.22) by (6).24 and
decrease the number of nonlinear terms. Analyze the number of nonlinear
terms before and after this substitution. What is the effect of this reformula-
tion?
6.5 Verify that (6.21), (6).24, and (6.25) imply (6.16).
6.6 For all the variables in the MINLP formulation for the multi-hop CRN
problem, what do we mean by core variables? What is the benefit of identi-
fying core variables in developing a solution?
6.7 Show how the values for the dependent variables tim , sijm , fij (l), and K
can be derived from the values for the core variables xijm and qijm .
6.8 When determining upper bounds for a subproblem, we need to perform
linear relaxation for two different nonlinear terms. Describe the technique we
used in linear relaxation of each nonlinear term. Can each technique in linear
relaxation be applied for both nonlinear terms?
6.9 In the local search, why do we initialize qijm at (qijm )L and increase it
upward? Discuss the case of initializing qijm at (qijm )U and decreasing it down-
ward.
6.10 When choosing a partitioning variable, why do we consider that the
x-variables are more important than the q-variables?
6.11 After partitioning a problem into two new subproblems on a xijm -
variable, how can the lower and upper bounds for the other core variables be
adjusted by using the first five constraints in Table 6.2?
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CHAPTER
7 Linear approximation
The great pleasure in life is doing what people say you cannot do.
Walter Bagehot
In contrast to LP, nonlinear optimization problems are usually hard and tedious
to solve. Therefore, one strategy for nonlinear programs is to replace all non-
linear terms by some linear approximation approach [11]. Depending on the
nature of the nonlinear program, the obtained LP problem may or may not
provide a feasible solution to the original nonlinear problem. In the case that
the LP solution is feasible, we may prove that it is a near-optimal solution
to the original problem. Even if it is infeasible, we may construct a feasible
solution via local search and further prove the constructed solution is near-
optimal.
In this chapter, we will explore such an approach to some nonlinear pro-
grams with special structures. This approach is also known as grid lineariza-
tion [88] or separable programming [11]: We consider the following problem:
Minimize f (x)
subject to gi (x) pi (1 i I )
hj (x) = qj (1 j J )
x 0,
where the objective function f (x) and constraint functions gi (x) are additively
separable, the constraint functions hj (x) are linear, and pi and qj are constants.
In other words, this implies that each f (x) and gi (x) can be expressed as a sum
of functions as follows:
148
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
()
()
= k1 k1 + k k
k1 + k = 1
k1 , k 0 .
() = k1 (k1 ) + k (k ).
It is quite apparent that as the number of grid points increases, the accuracy of
the approximation improves.
In general, for any [a, b], we can represent it as
m
= k k (7.1)
k=0
m
k = 1, (7.2)
k=0
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
where only two continuous k1 and k can be positive (when falls in the
kth segment). All other l , l = k 1, k, must be zero. The original function
() is approximated by
m
() = k (k ) . (7.3)
k=0
151 7.2 Case study: Renewable sensor networks with wireless energy transfer
Efforts at transferring power wirelessly can be dated back to the early 1900s
(long before wired electric power grid) when Nikola Tesla experimented with
large-scale wireless power distribution [156]. Due to its large electric fields,
which is undesirable for efficient energy transfer, Teslas invention was never
put into practice.
Since then, there was hardly any progress in wireless energy transfer for
many decades. In the early 1990s, the need for wireless power transfer re-
emerged when portable electronic devices became widely spread (see, e.g.,
[159]). The most well-known example is the electric toothbrush. However, due
to stringent requirements such as close contact, accurate alignment in charging
direction, and uninterrupted line of sight, most of the wireless power trans-
fer technologies at the time (based on inductive coupling) only found limited
applications.
Recently, wireless power transfer based on radio frequency (RF) between
850 MHz 950 MHz (with a center frequency of 915 MHz) has been explored
[125]. Under such radiative energy transfer technology, an RF transmitter
broadcasts radio waves in the 915 MHz ISM band and an RF receiver tunes
to the same frequency band to harvest radio power. However, it was found in
[93], [124], and [157] that a receiver operating under such radiative energy
transfer technology can only obtain about 45 mW power when it is 10 cm
away from the RF transmitter, with about 1% power transfer efficiency. A sim-
ilar experimental finding was reported in [66]. The technology is also sensitive
to obstructions between sources and devices, requires complicated tracking
mechanisms if relative positions change, and poses more stringent safety con-
cerns. Due to these issues, the potential of RF-based power transfer technology
is limited.
The foundation of this chapter is based on a recent breakthrough technol-
ogy by Kurs et al. [85], which was published in Science in 2007 and has
since caught worldwide attention. In [85], Kurs et al. experimentally demon-
strated that efficient nonradiative energy transfer was not only possible, but was
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Base station
also practical. They used two magnetic resonant objects having the same reso-
nant frequency to exchange energy efficiently, while dissipating relatively little
energy in extraneous off-resonant objects. They showed that efficient power
transfer implemented in this way can be nearly omnidirectional, irrespective
of the environment and even none-line-of-sight (NLOS). The power transfer
efficiency, however, decreases with distance. A highlight of their experiment
was to fully power a 60-W light bulb from a distance of two meters away, with
about 40% power transfer efficiency.
Since the first demo by Kurs et al. in 2007, there has been some rapid
advance on wireless energy transfer, particularly in the area of making it
portable. In particular, Kurs et al. launched a start-up company [171] and in
2009 they developed and demonstrated wireless energy transfer for portable
devices such as cell phones [49]. Note that the source coil remains sizable,
but the device coil is already portable (corresponding to our WCV and sensor
node, respectively).
With the recent establishment of Wireless Power Consortium [172] to set
the international standard for interoperable wireless charging, it is expected
that wireless power transfer will revolutionize how energy is replenished in the
near future.
Each sensor node consumes energy for data transmission and reception.
Denote pi the energy consumption rate at sensor node i N . In this chap-
ter, we use the energy consumption model in Section 2.3.1. We have
k=i
j =i
pi = fki + Cij fij + CiB fiB (i N ) , (7.9)
kN j N
where is the energy consumption for receiving a unit of data, Cij (or
CiB ) is the energy consumption for transmitting a unit of data from node
i to node j (or the base station B). Further, Cij = 1 + 2 Dij , where Dij
is the distance between nodes i and j , 1 is a distance-independent con-
stant term, 2 is a coefficient of the distance-dependent term, and is the
k=i
path-loss index. In the model, kN fki is the energy consumption rate
j =i
for reception, and j N Cij fij + CiB fiB is the energy consumption rate for
transmission.
To re-charge the battery at each sensor node, a mobile wireless charging
vehicle (WCV) is employed in the network. The WCV starts from a service
station (S), and the traveling speed of the WCV is V (in m/s). When it arrives
at a sensor node, say i, it will spend an amount of time i to charge the sensor
nodes battery wirelessly via wireless power transfer [85]. Denote U as the
energy transfer rate of the WCV. After i , the WCV leaves node i and travels
to the next node on its path. We assume that the WCV has sufficient energy to
charge all sensor nodes in the network.
After the WCV visits all the sensor nodes in the network, it will return to
its service station to be serviced (e.g., replacing or re-charging its battery) and
get ready for the next trip. We call this resting period vacation time, denoted
as vac . After this vacation, the WCV will go out for its next trip.
Denote as the time for a trip cycle of the WCV. A number of important
questions need to be addressed for such a network. First and foremost, we
would ask whether it is possible for each sensor node never to run out of its
energy. If this is possible, then the sensor network will have unlimited lifetime
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
and will never cease to be operational. Second, if the answer to the first ques-
tion is positive, then is there an optimal plan (including traveling path, stop
schedule) such that some useful objective can be maximized or minimized?
For example, in this chapter, we would like to maximize the percentage of
time in a cycle that the WCV spends on its vacation (i.e., vac
), or equivalently,
to minimize the percentage of time that the WCV is out in the field.
Table 7.1 lists the notation used in this chapter.
i1 D
k k+1
i1
ai = + + k . (7.10)
V
k=0 k=1
= P + vac + i , (7.11)
iN
where iN i is the total amount of time the WCV spends near all the sensor
nodes in the network for wireless energy transfer.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
ai Arrival time of the WCV at node i during the first renewable cycle
B Denotes the base station
Cij (or CiB ) Energy consumption for transmitting a unit of data from node i to
node j (or the base station B)
Dij (or DiS ) Distance from node i to node j (or the service station S)
DP The traveling distance during a renewable cycle under a traveling
path P
DTSP The minimal traveling distance during a renewable cycle
Emax Full battery capacity of sensor nodes
Emin Minimal energy reserve to keep a sensor node operational
Ei Starting energy of sensor node i in a renewable cycle with fully
re-charged battery
ei (t) Energy level of sensor node i at time t with fully re-charged battery
fij (or fiB ) Flow rate from sensor node i to node j (or the base station B)
h The reciprocal of
m The number of piecewise linear segments used to approximate
N The set of sensor nodes in the network
P The traveling path of the WCV
pi Energy consumption rate at sensor node i
Ri Data rate generated at sensor node i
S Denotes the service station
U Full charge rate of the WCV
ui Charge rate at sensor node i during the initial transient cycle
V Moving speed of the WCV
zi,k Binary variable indicating whether i falls within the kth segment
Path-loss index
1 A distance-independent constant term in energy consumption for
data transmission
2 A coefficient of the distance-dependent term in energy
consumption for data transmission
The power consumption coefficient for receiving data
Targeted approximation error, 0 < 1
i The ratio of the charging time at node i to the entire cycle time.
vac The ratio of the vacation time to the entire cycle time.
i An approximation of i2
i,k The weight of the grid point mk for
i
i The ith node traversed by the WCV along path P
Overall time spent during a renewable cycle
i Time span for the WCV to be present at the sensor node i for
re-charging battery
vac Vacation time during a renewable cycle
P Traveling time of the WCV during a renewable cycle under a
traveling path P
TSP Minimal traveling time of the WCV during a renewable cycle
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Base station
ei
Initial transient cycle First renewable cycle Second renewable cycle
Emax
Ei
Initial transient cycle
not shown here
Gi
(To be constructed later)
Emin t
0
0 ai ai + i 2 ai + ai + + i 3
pi = i U (i N ) . (7.12)
Note that when the WCV visits a node i at time ai during a renewable energy
cycle, it does not have to re-charge the sensor nodes battery to Emax . This is
illustrated in Fig. 7.4, where Gi denotes the starting energy of sensor node i in
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
a renewable cycle and gi (t) denotes the energy level at time t (dashed sawtooth
graph). During a cycle [, 2 ], we see that the energy level has only two slopes:
(i) a slope of pi when the WCV is not at this node (i.e., noncharging period),
and (ii) a slope of (U pi ) when the WCV is charging this node at a rate of U
(i.e., charging period). It is clear that gi (ai ) gi (t) gi (ai + i ), i.e., node
is energy level is lowest at time ai and is highest at time ai + i .
Also shown in Fig. 7.4 is another renewable energy cycle (in solid sawtooth
graph) where the battery energy is charged to Emax during a WCVs visit.
For this energy curve, denote Ei as the starting energy of node i in a renewable
be an optimal
cycle and ei (t) as the energy level at time t, respectively. Let Full
solution with fully re-charged battery in each cycle that maximizes the ratio of
the WCVs vacation time over the cycle time. Let be an optimal solution,
where there is no requirement on whether or not a nodes battery is fully re-
charged. Naturally, the optimal objective obtained by Full is no more than the
optimal objective obtained by due to the additional requirement (battery is
. Surprisingly, the following lemma shows that is
fully re-charged) in Full Full
equally good as in terms of maximizing the ratio of the WCVs vacation
time over the cycle time. Thus, for our optimization problem, it is sufficient to
consider a solution with fully re-charged battery.
Lemma 7.1
can achieve the same maximum ratio of vacation time to cycle
Solution Full
time as that for solution .
Proof. Our proof has two parts. (i) First, we show that the maximal ratio of
vacation time to cycle time achieved by solution is greater than or equal to
that achieved by solution Full . (ii) Second, we show the converse is also true,
i.e., the maximal ratio of vacation time to cycle time achieved by solution Full
is also greater than or equal to that achieved by solution . If both (i) and (ii)
hold, then the lemma is proved.
Since Full is an optimal solution with the additional requirement (fully
re-charged battery in each cycle), the maximal ratio of vacation time to the
cycle time obtained by Full is no more than that obtained by . Thus, (i)
holds.
We now prove (ii). Instead of considering optimal solution , we will prove
that any ratio achieved by a feasible solution can also be achieved by a fea-
sible fully re-charged solution . If this is true, in the special case that =
is an optimal solution, we have that the maximal ratio achieved by can also
be achieved by a feasible fully re-charged solution. Therefore, (ii) will hold.
The proof is based on construction. Suppose = (P , ai , Gi , fij , fiB , ,
i , P , vac , pi ) is a feasible solution to our problem. We construct = (P , ai ,
Ei , fij , fiB , , i , P , vac , pi ) by letting P = P , ai = ai , Ei = Emax + pi
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Based on Lemma 7.1, we will only consider a renewable cycle where each
node is fully re-charged when it is visited by the WCV. Since the energy level
at node i is at its lowest at time ai , to ensure the second requirement in Defini-
tion 7.1 we must have ei (ai ) = Ei (ai )pi Emin . Since for a renewable
cycle,
Lemma 7.2
A cycle is a renewable energy cycle if and only if constraints (7.11), (7.12),
and (7.14) are met for each sensor node i N .
Proof. The only if part of the lemma can be proved by showing that a renew-
able cycle meets (7.11), (7.12), and (7.14). This has already been shown in the
description of the renewable cycle.
We now prove the if part of the lemma, i.e., if (7.11), (7.12), and (7.14)
hold, then (i) and (ii) in Definition 7.1 will also hold, thus the cycle is a renew-
able energy cycle. Since (7.11) holds, the given cycle satisfies the time con-
straint. Constraint (7.12) ensures that the amount of energy charged to each
sensor node i during i is equal to the amount of energy consumed by sensor
node i in the cycle. So the energy level of each sensor node i at the end of the
cycle is the same as that at the beginning of the cycle. Therefore, requirement
(i) in Definition 7.1 is satisfied.
During the first renewable cycle, the lowest energy level at node i occurs at
time ai , which is
where the second equality holds by requirement (i), which we just proved, the
third equality holds by (7.13), the fourth equality holds by ei (ai + i ) = Emax
in a fully re-charged solution, and the last inequality holds by (7.14). Since
the lowest energy level of node i occurs at time ai and is still no less than
Emin , requirement (ii) in Definition 7.1 is met. The proof of the if part of the
lemma is complete.
Property 7.1
In an optimal solution, there exists at least one node in the network with its
battery energy dropping to Emin when the WCV arrives at this node.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Proof. The proof is based on contradiction (i.e., if this is not true, then we can
further increase the objective value, thus leading to contradiction).
Suppose there exists an optimal solution = (P , fij , fiB , , ,
i
, p ), where none of the nodes in the network has its energy
P , vac i
ever drops to Emin , i.e., ei (t) > Emin for all i N , t . Then we can
construct a new solution = (P , fij , fiB , , i , P , vac , pi ) by choosing
Emin
= miniN { E(max )p } 1 and letting P = P , fij = fij , fiB = fiB , =
i i
(1 + ) , i = (1 + )i , P = P , vac = vac + ( iN i ), and
pi = pi .
Now we show > 0. Since ei (t) > Emin for all i N , t , we have
ei (ai ) = Emax ( i )pi > Emin for all i N , i.e., miniN
In this section, we show that the WCV must move along the shortest Hamil-
tonian cycle in an optimal solution. This is formally stated in the following
theorem:
Theorem 7.1
In an optimal solution with the maximal vac
, the WCV must move along the
shortest Hamiltonian cycle that crosses all the sensor nodes and the service
station.
Proof. Theorem 7.1 can be proved by contradiction. That is, if there is an opti-
mal solution = (P , fij , fiB , , , , , p ), where the WCV does
i P vac i
not move along the shortest Hamiltonian cycle, then we can construct a new
solution = (P , fij , fiB , , i , P , vac , pi ), with the WCV moving along the
shortest Hamiltonian cycle and with an improved objective.
By assumption, P in does not follow the shortest Hamiltonian cycle.
The new solution is constructed as follows. Let P follows the shortest hamil-
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
and the solution for a renewable cycle becomes = (PTSP , fij , fiB , , i ,
TSP , vac , pi ). Since the optimal traveling path is determined, the solution can
be simplified as = (fij , fiB , , i , vac , pi ).
We note that the shortest Hamiltonian cycle may not be unique. Since any
shortest Hamiltonian cycle has the same total path distance and traveling time
TSP , the selection of a particular shortest Hamiltonian cycle does not affect
constraint (7.16), and yields the same optimal objective. This insight is for-
mally stated in the following corollary:
Corollary 7.1
Any shortest Hamiltonian cycle can achieve the same optimal objective.
We also note that to travel the shortest Hamiltonian cycle, there are two
(opposite) outgoing directions for the WCV to start from its home service sta-
tion. Since the proof of Theorem 7.1 is independent of the starting direction for
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
the WCV, either direction will yield an optimal solution with the same objec-
tive value, although some variables in each optimal solution will have different
values. We have the following corollary:
Corollary 7.2
The WCV can follow either direction to traverse the shortest Hamiltonian
cycle, both of which will achieve the same optimal objective. There exist
two optimal solutions corresponding to the two opposite directions, with
identical values of fij , fiB , , i , TSP , vac , pi , but different values of ai
(by (7.10)) and Ei (by (7.13)) due to difference in their respective renewable
cycles, where i, j N , i = j .
OPT
vac
Maximize
j =i
k=i
subject to fij + fiB fki = Ri (i N )
j N kN
k=i
j =i
fki + Cij fij
kN j N
In this problem, flow rates fij and fiB , time intervals , i , and vac , and power
consumption pi are optimization variables, and Ri , , Cij , CiB , U , Emax , Emin ,
and TSP are constants. This problem has both nonlinear objective ( vac ) and
nonlinear terms (pi and i pi ) in constraints (7.19) and (7.20).
Note that there are two possible outcomes for optimization problem OPT:
either an optimal solution exists or OPT is infeasible. There are several sce-
narios where the latter outcome may occur, e.g., (i) the energy charging rate of
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
WCV is too small or the energy consumption rate of a node is too large; (ii)
the time interval between WCVs visits at any node is too large. As a result,
some constraints in Problem OPT will not hold. These are physical limitations
for a WCV to achieve a renewable network lifetime for a WSN.
We note that in the above formulation, only constant TSP is related to the
shortest Hamiltonian cycle. Since this value does not depend on the traveling
direction along the Hamiltonian cycle, an optimal solution to Problem OPT
will work for either direction and yields a different renewable cycle for each
direction.
7.7.2 Reformulation
We first use the change-of-variable technique to simplify the formulation. For
the nonlinear objective vac
, we define
vac
vac = . (7.21)
For (7.18), we divide both sides by and have TSP 1 + vac + iN i = 1.
To remove the nonlinear terms 1 and i in the above equation, we define
i
i = (i N ) (7.22)
1
h= . (7.23)
Then (7.18) is reformulated as TSP h + vac + iN i = 1, or equivalently,
1 iN i vac
h= . (7.24)
TSP
Similarly, (7.19) and (7.20) can be reformulated (by dividing both sides
by ) as
pi = U i (i N ), (7.25)
(1 i ) pi (Emax Emin ) h (i N ) (7.26)
U TSP
vac 1 k i (1 i ) (i N ) .
Emax Emin
kN
OPT-R
Maximizevac
j =i
k=i
subject to fij +fiB fki = Ri (i N )
j N kN
k=i
j =i
fki + Cij fij + CiB fiB U i = 0 (i N ) (7.27)
kN j N
U TSP
vac 1 k i (1 i ) (i N ) (7.28)
Emax Emin
kN
fij , fiB 0, 0 i , vac 1 (i, j N , i = j ).
In this problem, fij , fiB , i , and vac are optimization variables, and Ri , ,
Cij , CiB , U , Emax , Emin , and TSP are constants. Algorithm 7.1 shows how to
obtain a solution to problem OPT once we obtain a solution to problem OPT-R.
After reformulation, the objective function and the constraints become lin-
ear except (7.28), where we have a second order i2 term, with 0 i 1.
In the next section, we present an efficient linear approximation technique
to approximate this second order nonlinear term (with performance guaran-
tee). This is the main technique that we wish to show in this chapter. Sub-
sequently, we develop an efficient near-optimal solution to our optimization
problem.
Remark 7.1
In our optimization problem, data routing and charging time are closely
coupled. We may want to de-couple routing from the charging problem
and require certain energy-efficient routing, e.g., the minimum energy
routing.1 However, minimum energy routing cannot guarantee optimal-
ity. This is because, to maximize vac , by (7.28), we need to minimize
U TSP
maxiN { kN k + Emax E min
i (1 i )}, i.e., minimize kN k +
U
maxiN { Emax E
TSP
i (1 i )}. But under minimum energy routing, we
min
k=i j =i
can only guarantee that iN ( kN fki + j N Cij fij + CiB fiB ) is
k=i
minimized. By the relationship in (7.27), minimizing iN ( kN fki +
j =i
j N Cij fij + CiB fiB ) is equivalent to minimizing iN i , which
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
U TSP
is only part of kN k + maxiN { Emax E min
i (1 i )}. Therefore,
minimum energy routing cannot guarantee the optimality of our problem.
This insight will be confirmed by our numerical examples in Section 7.9.
m
zik = 1 (7.29)
k=1
m
k
i = ik (7.30)
m
k=0
m
ik = 1 (7.31)
k=0
m
k2
i = ik (7.32)
m2
k=0
i0 zi1 (7.33)
ik zik + zi,k+1 (1 k < m) (7.34)
im zim , (7.35)
where zik is a binary variable indicating whether i falls within the kth segment
and ik [0, 1] is a weight associated with grid point mk . The setting of m will
determine the level of accuracy and will be studied in Section 7.7.4.
Since y = x 2 is a convex function, the piecewise linear approximation curve
(i , i ) lies above the curve (i , i2 ), 0 i 1. Thus, we have i i2 (see
Fig. 7.5). The following lemma characterizes the approximation error i i2
as a function of m.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
i
2i
1 1 3
0 4 i 2 4 1 X
Lemma 7.3
1
i i2 for i N .
4m2
where the first equality holds by (7.30), (7.32), and the fact that il = 0 for l =
k 1, k, the third equality holds by (7.31) and il = 0 for l = k 1, k, and the
last inequality holds by i,k 2i,k 14 when 0 i,k 1. This completes the
proof.
U TSP
vac 1 k (i i ) (i N ). (7.36)
Emax Emin
kN
OPT-L
Maximize vac
j =i
k=i
subject to fij + fiB fki = Ri (i N )
j N kN
k=i
j =i
fki + Cij fij + CiB fiB
kN j N
U i = 0 (i N )
U TSP
vac 1 k (i i ) (i N )
Emax Emin
kN
m
zik = 1 (i N )
k=1
m
k
i ik = 0 (i N )
m
k=0
m
ik = 1 (i N )
k=0
m
k2
i ik = 0 (i N )
m2
k=0
i0 zi1 0 (i N )
ik zik zi,k+1 0 (i N , 1 k < m)
im zim 0 (i N )
fij , fiB 0, 0 i , vac , i 1 (i, j N , i = j )
zik {0, 1} (i N , 1 k m)
0 ik 1 (i N , 0 k m),
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
where fij , fiB , i , vac , zik , ik , and i are variables, Ri , , Cij , CiB , U , Emax ,
Emin , and TSP are constants. The new formulation can be solved by a solver
such as CPLEX [31].
Construction of a feasible near-optimal solution The solution to problem
OPT-L is likely to be infeasible to problem OPT-R (and problem OPT). But
based on this solution, we can construct a feasible solution to problem OPT.
Suppose = (fij , fiB , i , vac , zik , ik , i ) is the solution to problem
OPT-L. By observing (fij , fiB , i , vac ), we find that it satisfies all con-
straints to problem OPT-R except (7.28). To construct a feasible solution =
(fij , fiB , i , vac ) to problem OPT-R, we let fij = fij , fiB = fiB , i = i .
For vac , in order to satisfy (7.28), we define
) ,
U TSP
vac = min 1 k i (1 i ) .
iN Emax Emin
kN
It is easy to verify that this newly constructed solution satisfies all the con-
straints for problem OPT-R. Once we have this solution to problem OPT-R, we
can easily find a solution to problem OPT via Algorithm 7.1.
Lemma 7.4
For the feasible solution with objective value vac , we have vac vac
U TSP
4(Emax Emin ) 1
m2
.
To prove Lemma 7.4, we need two intermediate results for vac and vac ,
which are stated in Lemmas 7.5 and 7.6, respectively.
Lemma 7.5
For the optimal solution to problem OPT-L, we have
U TSP
vac = 1 k (max max ) ,
Emax Emin
kN
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Y
1
max
2
max
0 1 max 12 3
4
1 X
4
2
Figure 7.6 An illustration of max and its approximation max .
where
max max{i }
iN
2
and max is the piecewise linear approximation of max (see Fig. 7.6).
U TSP
Proof. It is easy to see that in , vac = miniN {1 kN k Emax Emin
(i i )}. Thus, to prove this lemma, it is sufficient to show maxiN {i
i } = max max , i.e., i i max max for each i N .
We consider the following two cases:
Case 1: 0 < max 12 : Based on the definition of max , we have 0 < i
max 12 for each i N . We now show that i i is a nondecreasing func-
tion when 0 < i 12 . Then it follows that i i max max .
m i < m for a particular
Suppose i falls within the lth segment, i.e., l1 l
l. We have
(l 2 (l 1)2 ) mi l(l 1)
= i
% & m2
2l 1 l(l 1)
= 1 i + . (7.37)
m m2
When 0 < i 12 , we have l m+1 2 if m is odd and l 2 if m is even. Then
m
1 m 0 for both cases. Thus, when 0 < i 2 and i falls within the lth
2l1 1
j j
% &
2(m l + 1) 1 (m l + 1)(m l + 1 1)
= 1 j +
m m2
m + 2l 1 (m l + 1)(m l)
= (1 j ) +
m m2
% &
2l 1 (m +2lmm) + (m2 2lm+l 2 +ml)
2
= 1 j +
m m2
% &
2l 1 l2 l
= 1 j +
m m2
= j j .
Lemma 7.6
For the constructed solution to problem OPT-R, we have
U TSP
vac = 1 k max (1 max ).
Emax Emin
kN
max (1 max ).
Combining both cases, we have i (1 i ) max (1 max ). This completes
the proof.
Proof of Lemma 7.4. Denote vac as the objective value obtained by solution
to the relaxed linear problem OPT-L. Since problem OPT-L is a relaxation
, i.e., . Therefore,
of problem OPT-R, vac is an upper bound of vac vac vac
vac vac vac vac
- .
U TSP
= 1 k (max max )
Emax Emin
kN
- .
U TSP
1 k max (1 max )
Emax Emin
kN
U TSP
= (max max
2
)
Emax Emin
U TSP 1
,
4(Emax Emin ) m2
where the second equality holds by Lemmas 7.5 and 7.6, the fourth inequality
holds by Lemma 7.3. This completes the proof.
Based on Lemma 7.4, the following theorem shows how to set m such that
vac vac for a given target performance gap (0 < 1).
Theorem 7.2 /0 1
U TSP
For a given , 0 < 1, if m = 4(Emax Emin ) , then we have vac
vac .
Proof. Lemma 7.4 shows that the performance gap /0 is vac vac 1
U TSP U TSP
4(Emax Emin ) m2 . Therefore, if we set m =
1
4(Emax Emin )
0
U TSP
4(Emax Emin ) , then we have
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
U TSP 1
vac vac
4(Emax Emin ) m2
U TSP 4(Emax Emin )
4(Emax Emin ) U TSP
=.
Criterion 7.1
At each node i N , its initial transient cycle must meet the following criteria:
(i) ei (0) = Emax and ei ( ) = Ei ; and (ii) ei (t) Emin for t [0, ].
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
ei
Initial transient cycle First renewable cycle
Emax
Ei
Emin
t
0 ai ai + i ai ai + i 2
pi ai
ui = + pi , (7.38)
i
where ui is the charging rate at node i during the initial transient cycle and
ai is the arrival time of the WCV at node i in the initial transient cycle (see
Fig. 7.7).
Now we need to show ui U where U is the full charging rate. First, we
have
i1 D
k k+1
i1
i1 D
k k+1
i1
ai = + k = + k = ai , (7.39)
V V
k=0 k=1 k=0 k=1
where the second equality holds by P = P and i = i , and the last equality
follows from (7.10). Further, by (7.12), we have U i = pi = (2 )
pi (ai + i ) pi . It follows that
(ai ) pi (U pi ) i . (7.40)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Then, we have
pi ai pi (ai ) (U pi ) i
ui = + pi = + pi + pi = U ,
i i i
where the first equality follows from (7.38), the second equality follows from
(7.39), and the third inequality follows from (7.40).
For the newly constructed , we have the following theorem:
Theorem 7.3
The constructed is a feasible transient cycle.
Proof. To prove that is a feasible initial transient cycle, we need to show that
the newly constructed satisfies Criterion 7.1. By our assumption, ei (0) =
Emax . At time (ai + i ), we have
where the second equality follows from (7.41), and the fourth equality follows
from (7.13). Therefore, Criterion 7.1(i) is met.
To show ei (t) Emin for t [0, ], it is sufficient to show that ei (ai )
Emin and ei ( ) Emin , since these two time instances are the local mini-
mum for ei (t) during t [0, ]. We have ei (ai ) = ei (0) pi ai = Emax
pi (ai ) Ei pi (ai ) = ei (ai ) Emin . Also, by (7.42), ei ( ) =
Ei Emin . Hence, ei (t) Emin , for t [0, ].
In summary, meets all the criteria of a feasible initial transient cycle. This
completes the proof.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
7.9.2 Results
50-node network We first present complete results for the 50-node net-
work. Table 7.2 gives the location of each node and its data rate for a 50-
node network. The shortest Hamiltonian cycle is found by using the Concorde
solver [28] and is shown in Fig. 7.8. For this optimal cycle, DTSP = 5821 m
and TSP = 1164.2 s. For the target = 0.01, by Theorem 7.2, we have
23 4 23 4
U TSP 5 1164.2
m= = =4,
4(Emax Emin ) 4 0.01 (10800 540)
which is a small number. In our solution, the cycle time = 30.73 hours, the
vacation time vac = 26.82 hours, and the objective vac = 87.27%.
In Corollary 7.2, the WCV can follow either direction of the shortest Hamil-
tonian cycle while achieving the same objective value vac = 87.27%. Com-
paring the two solutions, the values for fij , fiB , , i , TSP , vac are identical,
while the values of ai and Ei are different. This finding can be verified by
our simulation results in Table 7.3 (counter clockwise direction) and Table 7.4
(clockwise direction). As an example, Figs. 7.9(a) and (b) show the energy
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 7.2 Location and data rate Ri for each node in a 50-node network.
Node index Location (m) Ri (kb/s) Node index Location (m) Ri (kb/s)
cycle behavior of a sensor node (the 32nd node) under the two opposite travel-
ing directions, respectively.
By Property 7.1, we find that there exists an energy bottleneck node in the
network with its energy dropping to Emin during a renewable energy cycle.
This property is confirmed in our numerical results. This bottleneck node is
the 48th node, whose energy behavior is shown in Fig. 7.10.
In Section 7.7.2, we showed that minimum energy routing may not be opti-
mal for our problem (see Remark 7.1). This point is confirmed by our numeri-
cal results. In Fig. 7.11, we show that data routing in our solution differs from
the minimum energy routing for the 50-node network.
100-node Network Table 7.5 gives the location of each node and its data rate
for a 100-node network. The shortest hamiltonian cycle is shown in Fig. 7.12.
For this optimal cycle, DTSP = 7687 m and TSP = 1537.4 s. For the target
= 0.01, by Theorem 7.2, we have
23 4 23 4
U TSP 5 1537.4
m= = = 5.
4(Emax Emin ) 4 0.01 (10800 540)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
500
Base station
Service station
X(m)
0
0 500 1000
The solution for the 100-node network includes the cycle time = 58.52 hours,
the vacation time vac = 50.30 hours, and the objective vac = 85.95%. Addi-
tional results are shown in Table 7.6.
7.11 Problems
Table 7.3 The case of counter clockwise traveling direction: Node visited along the
path, arrival time at each node, starting energy of each node in a renewable cycle,
and charging time at each node for the 50-node network.
Table 7.4 The case of clockwise traveling direction: Node visited along the path,
arrival time at each node, starting energy of each node in a renewable cycle, and
charging time at each node for the 50-node network.
7.5 How does Theorem 7.1 help us simplify the problem described in Sec-
tion 7.4? Assuming Theorem 7.1 is not provided, is there any change in prob-
lem OPT?
7.6 Give a formal proof of Corollary 7.1.
7.7 In Corollary 7.2, we claim that two opposite directions of the shortest
Hamiltonian cycle can achieve the same optimal objective. Explain why. In
the two optimal solutions corresponding to the two opposite directions, which
variables are changed and which are not? Justify your answer with the numer-
ical results in Section 7.9.
7.8 Property 7.1 shows that for an optimal solution, there exists at least one
bottleneck node in the network. Under a specific network, the WCV has two
opposite outgoing directions starting from its home service station (both follow
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Emin=540 t(s)
0
0 10752 11235 110628 121380 121863 221256 232008 232491 331884
Ei=8461
Emin=540 t(s)
0
0 3116 3599 110628 113744 114227 221256 224372 224855 331884
Ei=714
Emin=540 t(s)
0
0 1833 3925 110628 112461 114553 221256 223089 225181 331884
500
Base station
X(m)
0
0 500 1000
(a) Data routing in our solution.
Y(m)
1000 Sensor node
500
Base station
X(m)
0
0 500 1000
(b) Minimal energy routing.
Find out the shortest Hamiltonian cycle for this network instance. For the
optimal cycle, what is the total traveling distance DTSP and the total traveling
time TSP ?
7.10 What is the benefit of reformulating problem OPT to problem OPT-R?
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 7.5 Location and data rate Ri for each node in a 100-node network.
Node index Location (m) Ri (kb/s) Node index Location (m) Ri (kb/s)
Mobile WCV
500
Base station
X(m)
0
0 Service station 500 1000
Table 7.6 The case of clockwise traveling direction: Node visited along the path,
arrival time at each node, starting energy of each node in a renewable cycle, and
charging time at each node for the 100-node network.
7.11 Why do we put data routing as a part of the optimization problem? Why
not simply use the minimum energy routing?
7.12 What is the purpose of using piecewise linear approximation for the
quadratic terms i2 ? How do we ensure that such an approximation can guar-
antee (1 )-optimality?
7.13 After solving problem OPT-L, why is it necessary to construct a feasible
solution to problem OPT-R (and problem OPT)? Verify the feasibility of the
constructed solution (in Section 7.7.3).
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
()
()
0 1 1 3 1 X
4 2 4
Service station
Sensor node
50
X(m)
0
0 50 100
7.14 We consider a small WSN consisting of three nodes (see Fig. 7.15). The
length of the shortest Hamiltonian cycle is 120 m for this network instance. The
traveling speed of the WCV is 1.2 m/s. Each sensor node has a battery capacity
of 10 KJ and is fully charged initially. The minimum energy at a sensor node
battery for it to be operational is 500 J. The wireless energy transfer rate is 5
W. We are looking for a solution within 1% from optimum.
Answer the following questions:
(a) How can we set m, i.e., the number of segments that we use in the piecewise
linear approximation such that 99% optimality is achieved?
(b) Use piecewise linear approximation and derive a relaxed linear formulation
in the forms of problem OPT-L.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Sensor node
Table 7.8 Location and data rate Ri for each node in a 20-
node network.
Sensor node
500
Base station
X(m)
0
0 Service station 500 1000
Table 7.9 The case of clockwise traveling direction: Node visited along the path, arrival
time at each node, renewable cycle starting energy of each node, and charging time
at each node for the 20-node network.
Node Visited ai (s) Ei (J) i (s) Node Visited ai (s) Ei (J) i (s)
Along the Path Along the Path
hours, the vacation time vac = 28.48 hours, and the objective vac = 90.50%.
Table 7.9 shows nodes visited along the path of clockwise traveling direction,
arrival time at each node, renewable cycle starting energy of each node, and
charging time at each node for the 20-node network.
Assuming the WCV follows the clockwise traveling direction, answer the
following questions:
(a) Based on Table 7.9, how to determine the energy consumption rate pi at
sensor node i? In particular, what are p5 , p6 , and p14 ?
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
(b) In the initial transient cycle, the WCV re-charges sensor node i at the charg-
ing rate of ui . How to determine the value of ui for sensor node i? In particular,
what are u5 , u6 , and u14 ?
(c) Theorem 7.3 shows that the constructed initial cycle is a feasible transient
cycle. Demonstrate this point at the 18th node. Draw the energy behavior of
this specific node during the initial transient cycle and the first two renewable
cycles.
(d) Recall that a bottleneck node in the network is defined as a node with
its energy dropping to Emin during a renewable energy cycle. Which node is
the bottleneck node in this network? Justify your answer and draw the energy
behavior of the bottleneck node during the initial transient cycle plus the first
two renewable cycles.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CHAPTER
Do not follow where the path may lead. Go, instead, where there is no path and
leave a trail.
Ralph Waldo Emerson
Recall that Part II of this book focuses on methods for near-optimal and
approximation solutions. Specifically, Chapters 5 to 7 follow the OR optimiza-
tion approach to develop a (1 )-optimal solution. In this and the following
chapters, we will show how to develop the (1 )-optimal solution by follow-
ing the CS algorithm design approach.
We first give a brief overview of the so-called approximation algorithms.
Such algorithms are designed to offer solutions that can approximate the
unknown optimal solution according to certain benchmark performance crite-
ria. In particular, in wireless network research, the most popular approximation
algorithms can be classified as constant-factor approximation algorithms and
(1 )-optimal approximation algorithms.
191
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Both of the above two types of approximation algorithms are common in the
literature to solve wireless network problems. The constant-factor approxima-
tion algorithms would only be useful if c is close to 1. But unfortunately, many
results in the literature offer results that may be far away from 1. Although
such results may still be of theoretical value, they are hardly useful in practice
(in the sense of offering an accurate performance benchmark for the design of
distributed networking algorithms and protocols).
Our preference is toward the design of (1 )-optimal approximation
algorithms, which will be presented in a case study in this chapter. As
expected, such algorithms are intellectually challenging and require much
novelty in their design. Nevertheless, should we be able to design such an
algorithm, then both its theoretical significance and practical value would be
assured.
The case study in this chapter is a classic and fundamental problem on base
station placement in a wireless sensor network (WSN). The setting of this prob-
lem is similar to that in Chapter 2. The problem here is to find the optimal
location for the base station so that the network lifetime (until any sensor node
runs out of energy) is maximized.
We will show that this problem can be formulated into a nonlinear program-
ming (NLP), which is nonconvex and NP-hard in general [46]. One approach
is to employ the branch-and-bound framework in Chapter 5. But that solu-
tion approach is tedious and not elegant. Instead, we will design an elegant
approximation algorithm that can guarantee (1 )-optimal network lifetime
performance.
The (1 )-optimal approximation algorithm that we are going to present
is based on several techniques, which makes it possible to reduce an infinite
search space to a finite-element search space for the base station location. The
main idea of the approximation algorithm is to exploit a clever way of dis-
cretizing the cost parameter associated with energy consumption into a geo-
metric sequence with tight upper and lower bounds. As a result, we can divide
a continuous search space into a finite number of subareas. By further exploit-
ing the cost property of each subarea, we can represent each subarea with a
so-called fictitious cost point (FCP), which is an N -tuple cost vector with
each element representing an upper bound on the cost for a corresponding sen-
sor node in the network. Based on these ideas, we can successfully reduce
an infinite search space for the base station location into a finite number of
points for each of which we can solve an LP problem to find an achiev-
able network lifetime and data routing solution. By comparing the achievable
network lifetime obtained among all the FCPs, we show that the largest is
(1 )-optimal. We also show that placing the base station at any physical
point in the subarea corresponding to the best FCP is (1 )-optimal. We
analyze the complexity of the approximation algorithm and show that it is
polynomial.
The remainder of this chapter is organized as follows. Section 8.3 describes
the network model used in the case study and formally states the base station
placement problem. In Section 8.4, we show that for a given base station loca-
tion, the maximum achievable network lifetime and the corresponding optimal
routing can be found via a single LP problem. To search the best base sta-
tion location, in Section 8.5, we narrow down the search space to the small-
est enclosing disk (SED) that covers all the sensor nodes in the network. In
Section 8.6, we divide the continuous search space of the SED into a finite
number of subareas and represent each subarea by an FCP. Thus, we can find
the best FCP corresponding to the maximum network lifetime. We then show
that by placing the base station at any point in the subarea corresponding to
the best FCP, we obtain a (1 )-optimal network lifetime. In Section 8.7, we
summarize all the steps as an algorithm. We also give an example for illus-
tration. In Section 8.8, we prove the correctness of the algorithm and ana-
lyze its complexity. In Section 8.9, we give some additional numerical exam-
ples illustrating the efficacy of the algorithm. Section 8.10 summarizes this
chapter.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
where fki (b/s) is the incoming bit-rate received by sensor i from sensor k and
is a constant coefficient.
In this chapter, we assume that the interference from different transmissions
has been effectively avoided by appropriate MAC layer scheduling. For low bit
rate and deterministic traffic pattern considered in this chapter, a contention-
free MAC protocol is not hard to design and we omit its discussion in this
chapter. Table 8.1 lists the notation used in this chapter.
Symbol Definition
A The search space for the base station, which can be the smallest
enclosing disk to cover all sensor nodes
Am The mth subarea in the search space
B The base station
Cij Power consumption coefficient for transmitting data from sensor i
to sensor j
ciB (p) Power consumption coefficient for transmitting data from sensor i
to base station B at point p
CiBmin , C max Lower and upper bounds of c (p)
iB iB
C[h] The transmission cost for the hth circle
dij (or diB ) Distance between sensor i and sensor j (or base station B)
ei Initial energy at sensor i
fij (or fiB ) Data rate from sensor i to sensor j (or base station B)
Hi Total number of circles for discretization at sensor node i for a given
K Total number of circles for discretization for a given
M Total number of subareas for discretization for a given
N Set of sensor nodes in the network
N = |N |, number of sensor nodes in the network
OA The center of the smallest enclosing disk A
pm Fictitious cost point (FCP) representation for the mth subarea
popt The best base station location
p The best location among M FCPs
p A point in the subarea corresponding to p
ri Sensing data rate produced at sensor i
RA The radius of the smallest enclosing disk A
Tm Maximum achievable network lifetime by placing the base station at pm
Topt Optimal network lifetime achieved by placing the base station at popt
T = max{Tm : m = 1, 2, . . . , M}
T (1 )-optimal network lifetime achieved by p
Vij (or ViB ) Total data volume from sensor i to sensor j (or base station B)
(xi , yi ) Location of sensor node i
Path-loss index
1 , 2 Constant coefficients in transmission power modeling
Desired approximation error, 0 < 1
Power consumption coefficient for receiving data
opt The best flow routing solution when the base station is at popt
Flow routing solution when the base station is at p
Flow routing solution when the base station is at p
sensor node j and base station B, respectively. Then the flow balance for each
sensor node i is
k=i
j =i
fki + ri = fij + fiB ,
kN j N
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
i.e., the sum of total incoming flow rates plus the self-generated data rate is
equal to the sum of the total outgoing flow rates. The energy constraint for
each sensor node i is
k=i
j =i
fki T + Cij fij T + ciB (p) fiB T ei ,
kN j N
i.e., total consumed energy due to receiving and transmission over time T can-
not exceed its initial energy ei . By (8.2), we have
0
ciB (p) = 1 + 2 (xB xi ) + (yB yi )
2 2 ,
As discussed earlier, the maximum network lifetime depends on both the base
station location and data routing. To begin with, we show that for a given base
station location, we can find the maximum network lifetime and optimal rout-
ing via a single LP problem. Here, the objective function is the network lifetime
T and the constraints are given by (8.4) and (8.5). Multiplying both sides of
(8.4) by T and denoting
Vij = fij T and ViB = fiB T , (8.6)
where Vij (or ViB ) can be interpreted as the total data volume from sensor
node i to sensor node j (or base station B) over time T , we have the following
formulation:
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Maximize T
k=i
j =i
subject to Vki + ri T Vij ViB = 0 (i N )
kN j N
k=i
j =i
Vki + Cij Vij + ciB (p)ViB ei (i N )
kN j N
Note that for a given base station location, the ciB (p)-values are constants.
Therefore, the above formulation is in the form of an LP problem. Once we
solve the above LP problem, we can obtain an optimal routing solution for fij
V
and fiB by fij = Tij and fiB = VTiB .
Although for a given base station location we can find the corresponding max-
imum network lifetime via a single LP problem, it is not possible to examine
all (infinite) points in the two-dimensional plane and select the point having
the maximum network lifetime.
As a first step, we show that it is only necessary to consider points inside the
so-called smallest enclosing disk (SED) [167],1 which is a unique disk with
the smallest radius that contains all the N sensor nodes in the network and can
be found in O(N) time [108]. This is formally stated in the following lemma:
Lemma 8.1
To maximize network lifetime, the base station location must be within the
smallest enclosing disk A that covers all the N sensor nodes in the network.
Proof. The proof is based on contradiction. That is, if the base station location
is not in the SED, then its corresponding network lifetime cannot be maximum.
To see this, assume that the optimal base station location is at point p, which
is outside the SED A (see Fig. 8.1). Denote OA as the center of the SED. Let
q be the intersecting point between the line segment [p, OA ] and the circle
1 In fact, we can consider points in an even smaller area, i.e., the convex hull of all sensor nodes.
However, using the convex hull does not reduce the order of complexity of our algorithm. On
the other hand, the use of the SED simplifies the discussion.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
RA
of the SED. Then for any sensor node i (all in A), we have diq < dip . Con-
sequently, Ciq < Cip . As a result, we can save transmission energy for every
sensor node i N by relocating p to q. This saving in energy at each node
increases network lifetime, which shows that point p cannot be the optimal
location to maximize network lifetime.
We have thus narrowed down the search space for base station B from a
two-dimensional plane to the SED A. However, the number of points in A
remains infinite. It is tempting to partition A into small subareas (e.g., a grid-
like structure), A1 , A2 , . . ., up to say AM , i.e.,
M
A= Am .
m=1
OA RA
4
min
DiB = 0,
max
DiB = Di,OA + RA .
min
CiB = 1 , (8.7)
max
CiB = 1 + 2 (DiB ) = 1 + 2 (Di,OA + RA ) .
max
(8.8)
q2
4
corresponding to costs C[1], C[2], . . . , C[Hi ], which are defined with the fol-
lowing geometric property:
C[h] = CiB
min
(1 + )h = 1 (1 + )h (1 h Hi ). (8.9)
C[Hi ] CiB
max
.
For example, for node 4 in Fig. 8.2, we have H4 = 3, i.e., D[3] is the circle
centered at node 4 that will completely contain the disk. As a result, with sensor
node i as the center, we have a total of Hi circles, each with cost C[h], h =
1, 2, . . . , Hi .
The above partitioning of the SED A is with respect to a specific node i. We
now perform the above process for all sensor nodes. These intersecting circles
will cut disk A into a finite number of irregular subareas, with the boundaries
of each subarea being either an arc (with a center at some sensor node i and
some cost C[h], 1 h < Hi ) or an arc segment of the SED A. As an example,
the SED A in Fig. 8.3 is now cut into 28 irregular subareas.
We now claim that under this subarea partitioning technique, for any point
in a given subarea, its cost to each sensor node in the network can be tightly
bounded quantitatively. This is because, with respect to each sensor node i,
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Definition 8.1
Denote the FCP for subarea Am (m = 1, 2, . . . , M) as pm , which is repre-
sented by an N-tuple vector with its ith element (i = 1, 2, . . . , N) being an
upper bound on the cost for any point in subarea Am corresponding to the
ith sensor node in the network.
That is, the N -tuple cost vector for FCP pm is [c1B (pm ), c2B (pm ), . . .,
cNB (pm )], with the ith element ciB (pm ) being
ciB (pm ) = C[hi (Am )], (8.12)
where hi (Am ) is determined by (8.11).
As an example, the FCP for subarea with corner points {q1 , q2 , q3 } in
Fig. 8.3 can be represented by a 4-tuple cost vector [c1B (pm ), c2B (pm ),
c3B (pm ), c4B (pm )] = [C[2], C[3], C[2], C[3]], where the first component
C[2] represents an upper bound on the cost for any point in this subarea to
sensor node 1, the second component C[3] represents an upper bound on the
cost (which is loose here) for any point in this subarea to sensor node 2, and
so forth.
We emphasize that the reason we use the word fictitious is that an FCP pm
cannot be mapped to a physical point within the corresponding subarea Am .
This is because there does not exist a physical point in subarea Am that has its
costs to all the N sensor nodes equal (one by one) to the respective N -tuple cost
vector embodied by pm simultaneously. As an example, any point within the
dark subarea bounded by corner points q1 , q2 , and q3 cannot have its costs cor-
respond to the four sensor nodes in the network equal to the respective element
in [C[2], C[3], C[2], C[3]] simultaneously, where [C[2], C[3], C[2], C[3]] is
the cost vector for the FCP that represents this subarea.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
The following important property for FCP pm will be used in the proof of
the (1 )-optimality guarantee of the approximation algorithm.
Property 8.1
For any point p Am and the corresponding FCP pm , we have
ciB (pm ) (1 + )ciB (p).
Proof. By the definitions of FCP pm (see (8.12)) and C[h] (see (8.9)), we have
Example 8.1
We use a small three-node network to illustrate the steps of the approxima-
tion algorithm. The location, data rate, and initial energy for each sensor
are specified in Table 8.2, where the units of distance, rate, and energy are
all normalized. Also, we set = 2, 1 = 1, 2 = 0.5, and = 1 under the
normalized units. For illustration, we set the error bound to = 0.2.2
and
3 6
1 7 11
4 14
0.8 8
1 12
0.6
Y 15
9
0.4 2
5
13
0.2 16
10
0
0 0.2 0.4 0.6 0.8 1
X
Figure 8.4 The SED is partitioned into 16 subareas for the three-node example.
In this section, we give a formal proof that the solution obtained by Algo-
rithm 8.1 is (1 )-optimal and analyze its complexity.
Denote popt as the optimal location for base station placement, Topt and opt
as the corresponding maximum network lifetime and flow routing solution, all
of which are unknown.
Denote p as the best FCP among the M FCPs pm , m = 1, 2, . . . , M,
based on their network lifetime performance. Denote by T and the
2 This is used here to simplify the illustration of each step. In Section 8.9, we use = 0.05 for
all computations.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Theorem 8.1
T (1 )Topt .
Lemma 8.2
For any given base station location p and corresponding optimal routing
solution and achievable network lifetime T (obtained via the stated LP
formulation), denote Am as the subarea that contains p for a given . Then
for the corresponding FCP pm , its achievable network lifetime Tm is at least
(1 ) of T , i.e., Tm (1 )T .
Proof. Instead of considering the optimal routing solution for FCP pm , we use
the same routing on pm , which is clearly suboptimal. That is, denoting Tm
as the network lifetime for FCP pm under , we have Tm Tm . Then we only
need to show that Tm (1 )T .
To show Tm (1 )T , we compute the total consumed energy on node
i N under for FCP pm and at time (1 )T , which is
k=i
j =i
fki (1 )T + Cij fij (1 )T + ciB (pm )fiB (1 )T
kN j N
k=i
j =i
< fki T + Cij fij T + (1 + )ciB (p)fiB (1 )T
kN j N
k=i
j =i
< fki T + Cij fij T + ciB (p)fiB T ei .
kN j N
The first inequality holds via Property 8.1. The last inequality holds by the
energy constraint corresponding to the routing solution for the point p. Thus,
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
the network lifetime Tm for FCP pm under the routing solution is at least
(1 )T . Thus, we have Tm Tm (1 )T .
With Lemma 8.2, we are ready to prove Theorem 8.1 as follows:
Proof. Consider the special case of Lemma 8.2 where the given base station
location p is the optimal location popt , with the corresponding optimal data
routing solution opt and network lifetime Topt . Following the same approach
as in Lemma 8.2, we can find a corresponding subarea Am that contains the
point popt with the corresponding FCP pm . As a result, for FCP pm , we have
Tm (1 )Topt . Thus, for the best FCP p among all the FCPs, we have
T Tm (1 )Topt .
Theorem 8.1 guarantees that the best network lifetime among the M FCPs
is at least (1 ) of Topt . Now, consider a point p in the subarea represented
by the best FCP p . We have the following theorem:
Theorem 8.2
T (1 )Topt .
Proof. Denote T as the network lifetime for point p under the same routing
solution for FCP p . Since is a suboptimal routing for p , we have T
T . Thus, to show T (1 )Topt , we only need to show that T T
(1 )Topt , where the second inequality follows from Theorem 8.1.
To establish T T , we compute the total consumed energy on node i N
under for the point p at time T , which is given by
k=i
j =i
fki T + Cij fij T + ciB (p )fiB T
kN j N
k=i
j =i
fki T + Cij fij T + ciB (p )fiB T ei .
kN j N
The first inequality holds by (8.11) and (8.12). The second inequality holds by
the energy constraint on p under the routing solution . Thus, the network
lifetime T for location p under is at least T . As a result, the maximum
network lifetime T for location p is at least T T (1 )Topt .
The complexity of Algorithm 8.1 can be measured by the number of LPs that
need to be solved, which is equal to the total number of subareas M. Hence,
we compute M.
The boundary segments of each subarea are defined by either an arc centered
at some sensor node i (with some cost C[h], 1 h < Hi , with Hi being given
by (8.10)), or an arc of the disk A. Since there are Hi 1 circles radiating from
each sensor node i and one circle for the disk A, the total number of circles
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 8.3 Each nodes location, data generation rate, and initial energy for a ten-node
network.
is K = 1 + iN (Hi 1). The maximum number of subareas M that can be
obtained by K circles is upper bounded by
M K 2 K + 2. (8.13)
0.6
Y
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X
Table 8.4 Each nodes location, data generation rate, and initial energy for a 20-node
network.
1.80, 1.05, 1.05] has the maximum network lifetime T = 82.86 among all the
FCPs. Subsequently, we place the base station at a point in the corresponding
subarea, say at point (0.31, 0.79). The corresponding network lifetime is T =
82.91 > T , which is also at least 95% of the optimal value. The flow routing
solution is depicted in Fig. 8.6.
Y 0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X
8.11 Problems
(b) Apply your algorithm of Part (a) to an example having at least five ran-
domly deployed nodes in a two-dimensional area.
8.5 Our discretization procedure in Section 8.6.1 partitions the SED into a
number of irregular subareas. Why is such an irregular subarea division more
attractive than a regular grid-type subarea division?
8.6 Derive (8.10) for Hi .
8.7 What is a fictitious cost point and how is it represented? What is the
difference between a fictitious cost point and a physical point?
8.8 In Definition 8.1, why should we use upper bounds on transmission costs
to define a fictitious cost point?
8.9 Consider the same setting as in 8.1, except that the three nodes 1, 2, and 3
are instead located at (0, 0), (0.93, 0.25), (0.43, 0.75), respectively. Enumerate
all the fictitious cost points in terms of their 3-tuple vectors.
8.10 If the transmission power at each node is upper bounded by a constant
U (U < ), how would you revise Algorithm 8.1?
8.11 Prove that (8.13) is correct.
8.12 In this chapter, we only considered how to optimally locate a single
base station. Suppose now that we have two base stations and we want to
determine where to place each of them optimally so as to maximize the net-
work lifetime. How would you extend Algorithm 8.1 for this two base station
problem? Note that in this case, each sensor can split its data and transmit them
simultaneously to two base stations.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CHAPTER
9 Approximation algorithm
and its applications Part 2
9.1 Introduction
This chapter is a sequel to the last chapter. In the last chapter, we showed
how to design a (1 )-optimal approximation algorithm for a base station
placement problem. The final solution gives a point (or any point within the
corresponding fictitious cost point (FCP) subarea) for the base station. In this
chapter, we extend the base station placement problem to the next level, where
the base station is allowed to move around in the wireless sensor network
(WSN).
The benefits of using a mobile base station to prolong sensor network life-
time are easy to understand. Since the base station is the sink node for data col-
lected by all the sensor nodes in the network, the set of sensor nodes near the
base station would carry a considerable burden in relaying traffic from other
sensor nodes to the base station. By allowing the base station to be mobile,
we could alleviate the traffic burden from this fixed set of sensor nodes to
other sensor nodes in the network, and thus extend network lifetime. Further,
given new advances in unmanned autonomous vehicle (UAV) and customized
robotics, having an unmanned vehicle carrying a base station for sensor data
collection may not be too far from reality.
Although the potential benefit of using a mobile base station to prolong sen-
sor network lifetime is significant, the technical difficulty of this optimization
problem is enormous. There are two components that are tightly coupled in
211
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
this problem. First, the location of the base station is time-dependent, i.e., at
different time instances, the base station may be at different locations. Second,
the multi-hop flow routing appears to be dependent on both time and location
of the base station. As a result, an optimization problem with the objective of
maximizing network lifetime needs to consider both base station location and
flow routing, both of which are also time-dependent.
The first result shows that as far as the network lifetime objective is con-
cerned, we can transform the time-dependent problem to a location (space)-
dependent problem. In particular, we show that flow routing only depends
on the base station location, regardless of when the base station visits this
location. Further, the specific time instances for the base station to visit a
location is not important, as long as the total sojourn time for the base sta-
tion to be present at this location is the same. This result allows us to focus
on solving a location-dependent problem.
The second result shows that to obtain a (1 )-optimal solution to the
location-dependent problem, we only need to consider a finite set of points
within the smallest enclosing disk (SED) for the mobile base stations loca-
tion. This result follows the same approach in Section 8.6. Then we can find
the optimal sojourn time for the base station to stay at each FCP (as well
as the corresponding flow routing solution) such that the overall network
lifetime (i.e., sum of the sojourn times) is maximized via a single LP prob-
lem. We prove that the proposed solution can guarantee that the achieved
network lifetime is at least (1 ) of the maximum (unknown) network
lifetime.
k=i
j =i
gki (t) + ri = gij (t) + giB (t),
kN j N
i.e., for node i, the sum of total incoming flow rates plus self-generated data
rate is equal to the total outgoing flow rates at time t. Note that in our problem,
data generated at each node should be transmitted to the base station in real
time.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
General notation
A The movement region for the base station
Cij Transmission energy cost from sensor i to sensor j
ciB (t) Transmission energy cost from sensor i to base station B at time t
ciB (p) Transmission energy cost from sensor i to base station B at point p
ei Initial energy at sensor node i
fij (p) Flow rate from sensor i to sensor j when base station B is at point p
fiB (p) Flow rate from sensor i to base station B when B is at point p
gij (t) Flow rate from sensor i to sensor j at time t
giB (t) Flow rate from sensor i to base station B at time t
N The set of sensor nodes in the network
N = |N |, the number of sensor nodes in the network
p(s) The arrived point when base station B traverses a distance s along P
P The traveling path for base station B
ri Bit rate generated at sensor node i
s(t) Cumulative distance traversed by base station B up to time t
S The total traversed distance at the end of network lifetime
u(s) = v(t1(s)) when the base station traverses point p(s) at time t
U (s) Sojourn time for base station B at distance s
v(t) The
velocity of base station B at time t
w(p) = sZ(p) u(s) when base station B traverses point p but never dwells
W (p) Sojourn time for base station B at point p
(x, y)(t) Location of base station B at time t
(xi , yi ) Location of sensor node i
Z(p) Set of distance s with p(s) = p
Path-loss index, 2 4
1 , 2 Two constant terms in power consumption model for transmission
Power consumption coefficient for receiving data
i.e., total consumed energy due to reception and transmission over time T can-
not exceed its initial energy ei . We have
0
ciB (t) = 1 + 2 (x(t) xi )2 + (y(t) yi )2 ,
The difficulty in the problem formulation lies in that the base station location
(x, y)(t) and flow routing gij (t) and giB (t) are all functions of time. This adds
considerable difficulty in the optimization problem. In this section, we show
that as far as network lifetime performance is concerned, such dependency on
time can be relaxed. Specifically, we will show (Lemma 9.1) that flow routing
only needs to be dependent on the location of the base station and can be
independent of when the base station is present at this location. Further, as
long as the total sojourn time for the base station to be present at this location
is the same, the specific time instance (i.e., when) the base station visits this
location is not important (Lemma 9.2). These results effectively transform the
problem to a location-dependent problem.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Definition 9.1
A time-dependent solution consists of a network lifetime T , a path P =
{(x, y)(t) : t [0, T ]} for the base station, and a flow routing gij (t) and
giB (t) at time t, i, j N , i = j , where (x, y)(t), gij (t) and giB (t) are all
functions of time t.
For such a solution, denote v(t) as the base station velocity8at time t (and
t
thus v(t) is the base station speed at time t). Denote s(t) = =0 v( )d
as the distance traversed by the base station up to time t. Suppose the total
traversed distance at the end of network lifetime T is s(T ) = S. Then we have
s(t) [0, S] for t [0, T ]. Note that for a given path P, we can identify the
corresponding base station location for any s, which we denote as p(s). Denote
U (s) as the sojourn time at distance s. The base station may visit the same point
multiple times. Then multiple distances may correspond to the same point.
Denote Z(p) as the set of such distances that correspond to the same point p.
Then the total sojourn time at a point p is
W (p) = U (s). (9.1)
sZ(p)
Definition 9.2
A location-dependent solution consists of a path P for the base station,
W (p) at each point p P, flow routing fij (p) and fiB (p), i, j N , i =
j , when the base station is at point p, and a network lifetime T , where
W (p), fij (p), and fiB (p) are all functions of location p.
The following theorem shows that for the objective of network lifetime max-
imization, it is sufficient to consider location-dependent solutions.
Theorem 9.1
The optimal location-dependent solution can achieve the same maximum
network lifetime as the optimal time-dependent solution.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Lemma 9.1
Given a feasible time-dependent solution, we can construct a location-
dependent solution with the same network lifetime.
Proof. The proof is based on the following construction. For a given time-
dependent solution , it consists of a network lifetime T , a path P for the base
station, and a flow routing gij (t) and giB (t), i, j N , i = j . To construct a
location-dependent solution , we let the base station follow the same path
P and for each point p P, we compute W (p) by (9.1). For , we define
location-dependent flow rates fij (p) and fiB (p) for each point p P by the
average of gij (t) and giB (t) over all visits to p during [0, T ] as follows:
If the base station dwells at p at least once (with W (p) > 0), we define
8 (x,y)(t)=p
t[0,T ] gij (t)dt
fij (p) = , (9.2)
W (p)
8 (x,y)(t)=p
t[0,T ] giB (t)dt
fiB (p) = . (9.3)
W (p)
If the base station traverses p (maybe multiple times) but never dwells, then
there is a unique time corresponding to each s Z(p). Denote such time as
t (s). Define
1
u(s) = , (9.4)
v(t (s))
w(p) = u(s). (9.5)
sZ(p)
To show the data routing scheme with fij (p) and fiB (p) is feasible and
has the same network lifetime T , we need to prove that (i) when the base
station visits each point p P, flow conservation holds at this node, and (ii) at
time T , the energy consumption at each node is the same as that in solution .
Both (i) and (ii) can be intuitively explained by noting that fij (p) and fiB (p)
are defined by the average of gij (t) and giB (t), respectively.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
8 (x,y)(t)=p
The first equality holds by (9.2) and the fact that W (p) = t[0,T ] 1dt. The
third equality holds by the flow conservation in solution . The last equality
holds by (9.2) and (9.3).
If the base station traverses (but never dwells at) p, then we have the follow-
ing flow conservation:
k=i k=i
sZ(p) gki (t (s))u(s) sZ(p) ri u(s)
fki (p) + ri = +
w(p) w(p)
kN kN
! k=i "
sZ(p) kN gki (t (s)) + ri u(s)
=
w(p)
! "
k=i
sZ(p) kN gij (t (s)) + giB (t (s)) u(s)
=
w(p)
j =i
sZ(p) gij (t (s))u(s) sZ(p) giB (t (s))u(s)
= +
w(p) w(p)
j N
j =i
= fij (p) + fiB (p).
j N
The first equality holds by (9.6) and (9.5). The third equality holds by the flow
conservation in solution . The last equality holds by (9.6) and (9.7).
(ii) For energy consumption at time T , we want to show that the energy
consumption at each node i in the constructed location-dependent solution
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
U
(s)>0 7 U (s)=0
+ ciB (p(s))fiB (p(s))U (s) + ciB (p(s))fiB (p(s))u(s)ds
s[0,S] s[0,S]
k=i
7 T j =i
7 T 7 T
= gki (t)dt + Cij gij (t)dt + ciB (t)giB (t)dt,
kN 0 j N 0 0
where ciB (p) is the energy cost from sensor i to the base station B when the
base station is at point p. To show that the above equality holds, it is sufficient
to show that the following three equalities hold:
U
(s)>0 7 U (s)=0 7 T
fki (p(s))U (s) + fki (p(s))u(s)ds = gki (t)dt
s[0,S] s[0,S] 0
We now prove (9.10). The proofs for (9.8) and (9.9) are very similar (but
simpler) and are left as homework problems.
On the left-hand-side (LHS) of (9.10), we observe that the summation and
integration are over distances in [0, S]. Recall that Z(p) denotes the set of
total traversed distances when the base station visits p P. If the base station
visits p multiple times, then Z(p) has multiple elements. However, for each
s Z(p), ciB (p(s)) is the same since p(s) is the same point p. Further, based
on the definitions in (9.3) and (9.7) for fiB (), we have that fiB (p(s)) is also the
same for each s Z(p). Thus, for a point p P, we can group these distances
in Z(p) together in the summation as well as integration on the LHS of (9.10).
Now we select one distance from each group Z(p), p P, to represent this
group. In particular, we can select the smallest distance in Z(p). Denote Y(P)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
as the set of these representatives for each Z(p). Then Y(P) is a subset of
[0, S].
For the summation (first term) on the LHS of (9.10), we have
U
(s)>0
W (p(s))>0 U
(z)>0
ciB (p(s))fiB (p(s))U (s) = ciB (p(z))fiB (p(z))U (z)
s[0,S] sY(P) zZ(p(s))
W (p(s))>0 U
(z)>0
= ciB (p(s))fiB (p(s)) U (z)
sY(P) zZ(p(s))
W (p(s))>0
= ciB (p(s))fiB (p(s)) U (z)
sY(P) zZ(p(s))
W (p(s))>0
= ciB (p(s))fiB (p(s))W (p(s)).
sY(P)
(9.11)
The first equality holds by grouping those distances corresponding to the same
point p(s) that the base station dwells. The second equality holds by ciB (z) =
ciB (s) and fiB (p(z)) = fiB (p(s)) for each z Z(p(s)). The last equality holds
by (9.1).
Following the same token, for the integration (second term) on the LHS of
(9.10), it can be shown that
7 U (s)=0
ciB (p(s))fiB (p(s))u(s)ds
s[0,S]
7 W (p(s))=0
= ciB (p(s))fiB (p(s))w(p(s))ds. (9.12)
sY(P)
W (p(s))>0
LHS of (9.10) = ciB (p(s))fiB (p(s))W (p(s))
sY(P)
7 W (p(s))=0
+ ciB (p(s))fiB (p(s))w(p(s))ds
sY(P)
8 (x,y)(t)=p(s)
W (p(s))>0
t[0,T ] giB (t)dt
= ciB (p(s)) W (p(s))
W (p(s))
sY(P)
7
W (p(s))=0
sZ(p(s)) giB (t (s))u(s)
+ ciB (p(s)) w(p(s))ds
sY(P) w(p(s))
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
W (p(s))>0 7 (x,y)(t)=p(s)
= ciB (p(s)) giB (t)dt
sY(P) t[0,T ]
7 W (p(s))=0
+ ciB (p(s)) giB (t (s))u(s)ds, (9.13)
sY(P)
sZ(p(s))
7 v(t)>0 7 v(t)=0
RHS of (9.10) = ciB (t)giB (t)dt + ciB (t)giB (t)dt. (9.14)
t[0,T ] t[0,T ]
For the case of v(t) = 0 (i.e., the base station dwells at the current
point). Denote t1 (s) as the time when the base station has traversed a dis-
tance s and starts this dwelling period. Denote t2 (s) as the time when
the base station completes this dwelling period. Then U (s) > 0 and for
any time t during [t1 (s), t2 (s)], the base station dwells and v(t) = 0.
Thus,
8 v(t)=0
where the first equality holds as the integration t[0,T ] is limited to those
dwelling periods, each corresponding to a distance s [0, S] with U (s) > 0.
The second equality holds by ciB (t) = ciB (p(s)) for t [t1 (s), t2 (s)].
For the case of v(t) > 0 (i.e., the base station is traversing at the current
point p(s)). We have
7 v(t)>0 7 U (s)=0 1
ciB (t)giB (t)dt = ciB (p(s))giB (s) ds
t[0,T ] s[0,S] v(t (s))
7 U (s)=0
= ciB (p(s))giB (s)u(s)ds, (9.16)
s[0,S]
where giB (s) is the flow rate from sensor i to the base station B when
the base station is traversing at point p(s). The first equality holds by
ds = v(t (s))dt and the second equality holds by (9.4).
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
RHS of (9.10)
U
(s)>0 7 t2 (s) 7 U (s)=0
= ciB (p(s)) giB (t)dt + ciB (p(s))giB (s)u(s)ds.
s[0,S] t=t1 (s) s[0,S]
(9.17)
For the summation (first term) on the RHS of (9.17), we have
U
(s)>0 7 t2 (s)
ciB (p(s)) giB (t)dt
s[0,S] t=t1 (s)
W (p(s))>0 U
(z)>0 7 t2 (z)
= ciB (p(z)) giB (t)dt
sY(P) zZ(p(s)) t=t1 (z)
W (p(s))>0 U
(z)>0 7 t2 (z)
= ciB (p(s)) giB (t)dt
sY(P) zZ(p(s)) t=t1 (z)
W (p(s))>0 7 (x,y)(t)=p(s)
= ciB (p(s)) giB (t)dt, (9.18)
sY(P) t[0,T ]
where the first equality holds by grouping those distances corresponding to the
same point p(s) where the base station dwells. The second equality holds by
ciB (p(z)) = ciB (p(s)) for each z Z(p(s)).
Following the same token, for the integration (second term) on the RHS of
(9.17), it can be shown that
7 U (s)=0
ciB (p(s))giB (s)u(s)ds
s[0,S]
7 W (p(s))=0
= ciB (p(s)) giB (t (s))u(s)ds. (9.19)
sY(P)
sZ(p(s))
W (p(s))>0 7 (x,y)(t)=p(s)
RHS of (9.10) = ciB (p(s)) giB (t)dt
sY(P) t[0,T ]
7 W (p(s))=0
+ ciB (p(s)) giB (t (s))u(s)ds. (9.20)
sY(P)
sZ(p(s))
The following lemma further extends Lemma 9.1 and says that the ordering
and specific time instances for the base station to visit a particular point p is
not important.
Lemma 9.2
Under a location-dependent solution, as long as W (p) (and w(p)) at each
point p remains the same, the network lifetime T will remain unchanged
regardless of the ordering and frequency of the base stations presence at
each point.
k=i
j =i
Vki (pm ) + ri W (pm ) Vij (pm ) ViB (pm ) = 0
kN j N
(i N , 1 m M) (9.23)
M k=i
j =i
Vki (pm ) + Cij Vij (pm ) + ciB (pm ) ViB (pm ) ei
m=1 kN j N
(i N )
(9.24)
T , W (pm ), Vij (pm ), ViB (pm ) 0 (i, j N , i = j, 1 m M),
where (9.23) and (9.24) follow from (9.21) and (9.22), respectively. Once
we solve the above LP problem, we have W (pm ) for 1 m M. For each
point pm with W (pm ) > 0, we can obtain fij (pm ) and fiB (pm ) by fij (pm ) =
Vij (pm ) ViB (pm )
W (pm ) and fiB (pm ) = W (pm ) . Recall that for those points with W (pm ) = 0,
it means that the base station will not visit those points in this solution.
We summarize the result in this section with the following proposition:
Proposition 9.1
The C-MB problem can be solved via a single LP in polynomial time.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
The solution to the above LP problem yields the sojourn time for the base
station at each location pm , m = 1, 2, . . . , M, and the optimal flow routing
when the base station is at pm . So far, we assume that base station B can
move from one point to another in zero time. We will discuss how to relax this
assumption in Section 9.5.4.
Theorem 9.2
For a given > 0, define subareas Am and FCPs pm , m = 1, 2, . . . , M, as
in Section 9.5.2. Then we have TC-MB
(1 ) TU-MB .
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Network lifetime
solution.
(1e)T*UMB
CMB UMB
Lemma 9.3
Given a feasible solution U-MB to the U-MB problem with a network life-
time TU-MB , we can construct a solution C-MB to the C-MB problem with a
network lifetime TC-MB (1 ) TU-MB .
W (pm ) = (1 ) W (Am ),
and, for each point pm with W (pm ) > 0, set the flow routing when the base
station is at pm as
U(s)>0 7 U (s)=0
1
fij (pm ) = fij (p(s))U (s) + fij (p(s))u(s)ds ,
W (Am ) sSm
sSm
U(s)>0 7 U (s)=0
1
fiB (pm ) = fiB (p(s))U (s) + fiB (p(s))u(s)ds .
W (Am ) sSm
sSm
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
The details of the proof of Lemma 9.3 are similar to the proof of Lemma 9.1
and are left in a homework problem.
Lemma 9.3 is a powerful result. With this lemma, we are now ready to prove
Theorem 9.2.
Proof. Consider the special case of Lemma 9.3 where the given solution to
the U-MB problem is an optimal solution U-MB
with network lifetime TU-MB .
By Lemma 9.3, we can transform it into a solution to the C-MB problem with
a network lifetime at least (1 )TU-MB , i.e., there is a solution to the C-MB
problem on the FCPs with a network lifetime at least (1 )TU-MB . As a result,
the optimal solution C-MB to the C-MB problem must have a network lifetime
TC-MB
(1 )TU-MB .
Theorem 9.2 guarantees that the network lifetime obtained by the LP solu-
tion on the M FCPs is at least (1 ) of TU-MB . However, a FCP may not be
mapped to a physical point, which is required in the final solution. In the fol-
lowing theorem, we show how to construct a solution with each point being
physically realizable. Further, the network lifetime for this constructed solu-
tion is greater than or equal to the maximum network lifetime for the C-MB
problem, i.e., TU-MB TC-MB . As a result, this new solution is (1 )-optimal.
Theorem 9.3
For a given > 0, define subareas Am and FCPs pm , m = 1, 2, . . . , M,
as discussed in Section 9.5.2. Given an optimal solution C-MB on these
M FCPs with W (pm ), fij (pm ), fiB (pm ), and a network lifetime TC-MB ,a
(1 )-optimal solution U-MB to the U-MB problem can be constructed
by having the base station stay in Am for
W (Am ) = W (pm ) (9.26)
amount of time and by having a corresponding flow routing for any physical
point p Am as
fij (p) = fij (pm ), (9.27)
fiB (p) = fiB (pm ). (9.28)
In Theorem 9.3, note that in the constructed solution to the U-MB problem,
as long as the base station is within Am (any point in this subarea), the flow
routing is the same.
Proof. We will show that the constructed solution U-MB is feasible, i.e., (i)
flow conservation holds at any point, and (ii) the network lifetime of U-MB is at
least TC-MB
. Based on (i) and (ii), as well as the fact that TC-MB
(1 )TU-MB
(see Theorem 9.2), we know that TU-MB must be at least (1 )TU-MB , i.e.,
U-MB is (1 )-optimal.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
(i) For flow conservation, when the base station location p is in Am , we have
k=i
k=i
fki (p) + ri = fki (pm ) + ri
kN kN
j =i
= fij (pm ) + fiB (pm )
j N
j =i
= fij (p) + fiB (p).
j N
The first equality holds by (9.27). The second equality holds by the flow
conservation in solution C-MB . The third equality holds by (9.27) and
(9.28). Thus, solution U-MB is feasible.
(ii) The total energy consumption at node i N by time TC-MB in solution
U-MB is
k=i
U(s)>0 7 U (s)=0
fki (p(s))U (s) + fki (p(s))u(s)ds
kN s[0,S] s[0,S]
j =i
U
(s)=0 7 U (s)=0
+ Cij fij (p(s))U (s) + fij (p(s))u(s)ds
j N s[0,S] s[0,S]
U
(s)>0 7 U (s)=0
+ ciB (p(s))fiB (p(s))U (s)+ ciB (p(s))fiB (p(s))u(s)ds.
s[0,S] s[0,S]
(9.29)
For the last two terms in (9.29), we have
U
(s)>0 7 U (s)=0
ciB (p(s))fiB (p(s))U (s) + ciB (p(s))fiB (p(s))u(s)ds
s[0,S] s[0,S]
M U
(s)>0 7 U (s)=0
= ciB (p(s))fiB (p(s))U (s) + ciB (p(s))fiB (p(s))u(s)ds
m=1 sSm sSm
M U
(s)>0 7 U (s)=0
(p )U (s) +
ciB (pm )fiB (p )u(s)ds
ciB (pm )fiB
m m
m=1 sSm sSm
M U
(s)>0 7 U (s)=0
= (p )
ciB (pm )fiB U (s) + u(s)ds
m
m=1 sSm sSm
M
= (p )W (A )
ciB (pm )fiB m m
m=1
M
= (p )W (p )
ciB (pm )fiB (9.30)
m m
m=1
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
for i N . The second inequality holds by ciB (p(s)) ciB (pm ) in Prop-
erty 8.1 and (9.28). The fourth equality holds by (9.25). The last equality
holds by (9.26).
M
+ ciB (pm )fiB (pm )W (pm ) ei ,
m=1
where the inequality holds by the energy constraint in solution C-MB . Thus,
the network lifetime of solution U-MB is at least TC-MB (1 )TU-MB . This
completes the proof.
Table 9.2 Sensor location, data rate, and initial energy of the example
sensor network.
0.8
2
0.6
OA
RA
0.4
3
4
0.2
0
0 0.2 0.4 0.6 0.8 1 X
sensor node i N and one circle for disk A, the total number of subareas
M obtained through the intersection of these circles is upper bounded by
O([1 + N i=1 (Hi 1)] ) = O((N/) ). Thus, the LP problem in Step 5 has
2 2
Example 9.1
To illustrate the steps in Algorithm 9.1, we solve a small four-node WSN
problem as an example. The location, data rate, and initial energy for each
sensor node are shown in Table 9.2, where the units of distance, rate, and
energy are all normalized with appropriate dimensions. We use = 2 in
this example and set 1 = 1, 2 = 0.5 and = 1 under normalized units.
For illustration, we set = 0.2.1
In Step 1, we first identify the SED A with origin OA = (0.60, 0.55) and
radius RA = 0.53 (see Fig. 9.2). Then we have D1,OA = 0.53, D2,OA =
0.21, D3,OA = 0.25, and D4,OA = 0.53. We then find the lower and upper
bounds of ciB for each node i as follows:
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
min
CiB = 1 = 1,
max
CiB = 1 + 2 (Di,OA + RA ) .
Thus, we have
max
C1B = 1 + 0.5 (0.53 + 0.53)2 = 1.56,
max
C2B = 1 + 0.5 (0.21 + 0.53)2 = 1.27,
max
C3B = 1 + 0.5 (0.25 + 0.53)2 = 1.30,
max
C4B = 1 + 0.5 (0.53 + 0.53)2 = 1.56.
In Step 2, for = 0.2, since
2 2 4
ln(1 + 1 (Di,OA + RA ) )
Hi = ,
ln(1 + )
we have
2 4
ln(1 + 1 (0.53 + 0.53) )
0.5 2
H1 = = 3,
ln(1 + 0.2)
2 4
ln(1 + 1 (0.21 + 0.53) )
0.5 2
H2 = = 2,
ln(1 + 0.2)
2 4
ln(1 + 1 (0.25 + 0.53) )
0.5 2
H3 = = 2,
ln(1 + 0.2)
2 4
ln(1 + 1 (0.53 + 0.53) )
0.5 2
H4 = = 3,
ln(1 + 0.2)
and
C[1] = 1 (1 + ) = 1 (1 + 0.2) = 1.20,
C[2] = 1 (1 + )2 = 1 (1 + 0.2)2 = 1.44,
C[3] = 1 (1 + )3 = 1 (1 + 0.2)3 = 1.73.
In Step 3, we draw a sequence of circles centered at each node i, 1 i
4, and with cost C[h], 1 h < Hi , to divide the SED A into 16 subareas
A1 , A2 , . . . , A16 (see Fig. 9.3).
In Step 4, we define a FCP pm for each subarea Am , 1
m 16. For example, for FCP p1 , we define the 4-tuple cost vec-
tor as [c1B (pm ), c2B (pm ), c3B (pm ), c4B (pm )] = [C[1], C[1], C[2], C[3]] =
[1.20, 1.20, 1.44, 1.73].
In Step 5, we obtain an optimal solution C-MB to C-MB problem on
these 16 FCPs by the LP approach discussed in Section 9.5.1. We obtain the
network lifetime TC-MB = 247.76, W (p7 ) = 144.22, W (p12 ) = 82.50,
W (p16 ) = 21.04, and for all other 13 FCPs, we have W (pm ) = 0 (mean-
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
ing the base station will not visit these 13 subareas). When the base station
is at FCP p7 , the routing is f1B (p ) = 0.60, f (p ) = 0.51, f (p ) =
7 2B 7 23 7
0.49, f3B (p7 ) = 1.29, and f4B (p7 ) = 0.40. When the base station is at FCP
p12 , the routing is f12 (p ) = 0.60, f (p ) = 1.60, f (p ) = 0.80,
12 2B 12 3B 12
and f4B (p12 ) = 0.40. When the base station is at FCP p16 , the routing is
(p ) = 0.60, f (p ) = 1.60, f (p ) = 2.40, and f (p ) = 2.80.
f12 16 23 16 34 16 4B 16
1
3
1 5
9
6
0.8 10
11
15
13
2
0.6 4
7
0.4
12 16
0.2
8
14
0
0 0.2 0.4 0.6 0.8 1 X
9.5.4 Discussions
We now discuss the design of a path P based on W (pm ) values. Such a
path is certainly not unique. In 9.1, the base station can move from sub-
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 9.3 Each nodes location, data generation rate and initial energy for a ten-node
network.
Location Data rate Initial energy Location Data rate Initial energy
area 7 to 12 and to 16 (denote as (A7 , A12 , A16 )) or, another path can be
(A16 , A12 , A7 ). Note that for any path, as long as the total sojourn time at
each subarea Am is W (pm ), the achieved network lifetime is (1 )-optimal.
Thus, all of these paths are equally good under network lifetime objective.
It may be arguable that one path is better than another under some other
objective, e.g., minimizing the total traveled distance. However, such objec-
tive can be formulated as a separate problem and its discussion is beyond the
scope of this chapter.
Along a path P, it is possible that one subarea and the next subarea that the
base station visits are not adjacent. We argue that the traveling time between
two subareas (e.g., minutes) is likely on a much smaller time scale than net-
work lifetime (e.g., months). It can be shown that if buffering is available at
sensor nodes when base station is in transition from one subarea to the next
subarea, then the (1 )-optimal network lifetime can still be achieved. In
this case, a node only needs to slightly delay its transmission until the base
station arrives at the next subarea and then empties the buffer with a higher
rate for a brief period of time.
Am (x, y) W (Am )
0.6
Y
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X
Table 9.5 Each nodes location, data generation rate and initial energy for a 20-node
network.
Location Data rate Initial energy Location Data rate Initial energy
(0.52, 0.02) 0.6 480 (0.29, 0.14) 0.6 120
(0.74, 0.76) 0.3 310 (0.05, 0.99) 0.4 60
(0.95, 0.03) 0.8 150 (0.84, 0.06) 1.0 180
(0.53, 0.63) 0.6 220 (0.99, 0.37) 0.4 340
(0.58, 1.00) 0.4 230 (0.73, 0.67) 0.8 220
(0.48, 0.84) 0.7 160 (0.53, 0.27) 0.5 380
(0.17, 0.83) 0.1 380 (0.57, 0.05) 0.7 250
(0.73, 0.39) 0.1 500 (0.88, 0.84) 0.2 240
(0.36, 0.98) 0.1 430 (0.26, 0.12) 0.9 440
(0.76, 0.02) 0.7 500 (0.71, 0.21) 0.3 70
0.6
Y
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X
The network setting for a small 20-node network (with location, data rate,
and initial energy for each of the 20 sensor nodes) is given in Table 9.5. By
applying Algorithm 9.1, we obtain a (1 )-optimal network lifetime 144.23.
Again, we use a star to represent the subarea that base station will visit in
the solution. For this particular 20-node network setting, we have six subareas
(see Fig. 9.6) that the base station will visit in the final solution, with the cor-
responding sojourn time in each subarea shown in Table 9.6. Again, we show
a possible path for the 20-node network in Fig. 9.7.
The network setting for the 50-node network (with location, data rate, and
initial energy for each of the 50 sensor nodes) is given in Table 9.8. By apply-
ing Algorithm 9.1, we obtain a (1 )-optimal network lifetime 122.30. In
Table 9.7, we have eight subareas (see Fig. 9.8) that the base station will visit
in the (1 )-optimal solution, as well as the sojourn time for the base station
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Am (x, y) W (Am )
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
in each of these eight subareas. Fig. 9.9 shows a possible path for the 50-node
network.
Finally, we consider a 100-node network shown in Table 9.9. We omit
to list the each nodes coordinates, data rate, and initial energy to conserve
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Am (x, y) W (Am )
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Table 9.8 Each nodes location, data generation rate and initial energy for a 50-node
network.
Location Data rate Initial energy Location Data rate Initial energy
Table 9.9 Each nodes location, data generation rate and initial energy for a 100-
node network.
Location Data rate Initial energy Location Data rate Initial energy
Location Data rate Initial energy Location Data rate Initial energy
space. They are all randomly generated as we described early in this sec-
tion. By applying Algorithm 9.1, we obtain a (1 )-optimal network lifetime
149.45. For this particular 100-node network setting, we have 12 subareas (see
Fig. 9.10) that the base station will visit, with the corresponding sojourn time in
each subarea shown in Table 9.10. Fig. 9.11 shows a possible path for the 100-
node network.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Am (x, y) W (Am )
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
This chapter is a sequel to the last chapter. Again, our interest is in the design
of a (1 )-optimal approximation algorithm. But the problem is much harder
than that in the last chapter. By allowing the base station to be mobile, both the
location of the base station and the multi-hop flow routing in the network are
time-dependent.
To address this problem, we showed that as far as network lifetime objec-
tive is concerned, we can transform the time-dependent problem to a location
(space)-dependent problem. In particular, we showed that flow routing only
depends on the base station location, regardless of when the base station visits
this location. Further, the specific time instances for the base station to visit a
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
location are not important, as long as the total sojourn time for the base station
to be present at this location is the same. This result allowed us to focus on
solving a location-dependent problem.
Based on the above result, we further showed that to obtain a (1 )-
optimal solution to the location-dependent problem, we only need to consider
a finite set of points within the SED for the mobile base stations location.
Here, we followed the same approach as that in Section 8.6, i.e., discretization
of energy cost through a geometric sequence, division of a disk into a finite
number of subareas, and representation of each subarea with a FCP. Then we
can find the optimal sojourn time for the base station to stay at each FCP (as
well as the corresponding flow routing solution) so that the overall network
lifetime (i.e., sum of the sojourn times) is maximized via a single LP problem.
We proved that the proposed solution can guarantee that the achieved network
lifetime is at least (1 ) of the maximum (unknown) network lifetime.
This chapter offers some excellent examples on how to transform a prob-
lem from time domain to space domain and how to prove results through
construction. Students are encouraged to gain a deep understanding of these
techniques, which should transcend to other optimization problems in wireless
networks.
9.8 Problems
9.1 What is the benefit of using mobile base station? Compared to the static
base station problem in the last chapter, what are the challenges in employing
a mobile base station?
9.2 Show that the formulation for the optimization problem in Section 9.3 is
nonconvex.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
9.3 For the optimization problem formulated in Section 9.3, can we dis-
cretize time t to obtain an approximation algorithm?
9.4 Describe the three key techniques used in the design of the approxima-
tion algorithm in the chapter.
9.5 What is the relationship between a time-dependent solution and a
location-dependent solution?
9.6 Can we use Eqs. (9.2) and (9.3) to define fij (p) and fiB (p) for every
point p P? Why? Can we use Eqs. (9.6) and (9.7) to define fij (p) and
fiB (p) for every point p P? Why?
9.7 Prove Eq. (9.8).
9.8 Prove Lemma 9.2.
9.9 Complete the proof of Lemma 9.3.
9.10 In Section 9.5.4, it is said that the design of a path P based on W (pm )
values is not unique. We may use an additional objective to select the best path,
e.g., minimizing the total traveled distance. Discuss other possible objective(s)
for path selection.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
PART
CHAPTER
Remember, happiness doesnt depend upon who you are or what you have; it
depends solely upon what you think.
Dale Carnegie
245
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
more integer variables based on the closeness in the relaxed solution to certain
integer values. Instead of determining all integer variable values via a single
relaxation, we can fix only one or a few integer variables in each iteration. For
the remaining (unfixed) integer variables, we can solve a new relaxation (with
some integer variables values being already fixed) and then fix one or more
integer variables. This SF procedure terminates after we fix all the integer vari-
ables. The values of other variables in the original problem can be obtained by
solving the resulting LP problem.
Unlike most techniques in other chapters, SF is a heuristic procedure. To
measure its performance, we can compare its solution value to some perfor-
mance bound, e.g., a lower bound for a minimization problem, or an upper
bound for a maximization problem. Such a bound can be obtained via the ini-
tial relaxation to the original problem. Note that the optimal objective value lies
between this bound and the solution obtained by the SF algorithm. Therefore,
if we can show (e.g., via a large number of simulations) that the SF solution
has a value very close to the computed relaxation-based bound, then we can
claim that the SF solution must be even closer to the optimum, thus validating
its performance.
In the case study of this chapter, we show how SF can be employed to solve
a spectrum sharing problem in a multi-hop cognitive radio network (CRN).
Consider a multi-hop CRN. For such a network, each node senses a set of
spectrum bands that it can use for communication. Due to the unequal size
of spectrum bands, it may be necessary to further divide a large band into
subbands (likely of unequal size) to schedule transmission and reception.
There are many fundamental problems that can be posed for such a wireless
network in the context of rates and capacity. In this chapter, we consider the
following problem. Suppose there exists a set of user sessions in the network
that is characterized by several sourcedestination pairs, each having a certain
rate requirement. Then, how can we perform spectrum allocation, scheduling
and interference avoidance, and multi-hop flow routing such that the required
network-wide radio spectrum resource is minimized?
To formulate the problem mathematically, we characterize the structure
and constraints of a CRN at multiple layers. Special attention is given to
the modeling of spectrum sharing and unequal (nonuniform) subband divi-
sion, scheduling and interference modeling, and multi-hop routing. We for-
mulate an optimization problem with the objective of minimizing the required
network-wide radio spectrum resource for a set of session rate requirements.
Since such a problem formulation is a mixed-integer nonlinear programming
(MINLP), which is NP-hard in general [46], we aim to derive a near-optimal
solution.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Symbol Definition
More formally, we model the union of the available frequency bands among
all the nodes in the network as a set of M unequally sized bands (see Fig. 10.1).
Denote M as the set of these bands and let Mi M be the set of available
bands at node i N , which is likely to be different from that at another node,
say j N , i.e., possibly, Mi = Mj . For example, at node i, Mi may consist
of bands 1, 3, and 5, whereas at node j , Mj may consist of bands 1, 4, and
6. Denote W (m) as the bandwidth of band m M. For more flexible and effi-
cient bandwidth allocation and to overcome the disparity in the bandwidth size
among the spectrum bands, we assume that band m can be further divided into
up to K (m) subbands, each of which may be of unequal bandwidth. Denote
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
(m)
u(m,1)W (m) u(m,2)W (m) u(m,K )
W (m)
(m)
(m,1) (m,2) (m,K )
u(m,k) as the fraction of bandwidth for the k-th subband in band m, which is
part of our set of optimization decision variables. Then we have
(m)
K
u(m,k) = 1.
k=1
Note that some u(m,k) -variables can be 0 in the final optimization solution,
in which case we will have a fewer number of subbands than K (m) . As an
example, Fig. 10.1 shows M bands in the network, and for a specific band m,
it displays a further division into K (m) subbands. Thus, the M bands in the
network are effectively divided into M m=1 K
(m) subbands, each of which may
be of a different size.
Transmission range and interference range We assume that the power
spectral density from the transmitter of a CR node is . In this chapter, we
assume that all nodes use the same power density for transmission. A widely
used model for power propagation gain is [56]
where is an antenna related constant, is the path-loss index, and dij is the
distance between nodes i and j . We assume that data transmission is successful
only if the received power spectral density at the receiver exceeds a threshold
T . Likewise, we assume that an interference will become nonnegligible only
if it produces a power spectral density over a threshold of I at a receiver.
Based on the threshold T , the transmission range for a node is thus given
by RT = (/T )1/ , which is derived from (RT ) = T . Similarly,
based on the interference threshold I (< T ), the interference range for a node
is given by RI = (/I )1/ . Because I < T , we have RI > RT .
Scheduling and interference constraints Scheduling can be done either in
time domain or frequency domain. In this chapter, we consider a frequency
domain subband assignment, i.e., how to assign subbands at a node for trans-
mission and reception. A feasible scheduling on frequency bands must ensure
that there is no interference at the same node and among the neighboring
nodes.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
For a node i N and a band m Mi , denote Tim as the set of nodes that
can use band m and are within the transmission range to node i, i.e.,
Tim = {j : dij RT , j = i, m Mj }.
We assume that node i cannot transmit to multiple nodes on the same frequency
sub-band. We have
(m,k)
xiq 1. (10.2)
qTim
For a frequency subband (m, k), if node i uses this subband for transmitting
data to a node j Tim , then any other node that can produce interference on
node j should not use this subband. Note that the so-called hidden terminal
problem is a special case under this constraint. To model this constraint, we
denote Ijm as the set of nodes that can produce interference at node j on band
m, i.e.,
The physical meaning of Tpm = in the above definition is that node p may
use band m for a valid transmission to some node in Tpm , which then causes
interference to node j . Thus, we have
xij(m,k) + (m,k)
xpq 1(p Ijm , p = i). (10.3)
qTpm
RI
6
5 2
RI
1 3
(to node 6) as stated in (10.3). That is, both nodes 3 and 5 may use the same
subband for transmission.
Flow routing At the network level, a source node may need a number of
relay nodes to route the data stream toward its destination node. Clearly, a route
having only a single path may be overly restrictive and may not be able to take
advantage of load balancing. A set of paths (with flow splitting) is more flexible
to route the traffic from a source node to its destination. Mathematically, this
can be modeled as follows. Denote fij (l) as the data rate on link i j that is
attributed to session l, where i N , j mMi Tim , and l L. To simplify
the notation, let Ti = mMi Tim . If node i is the source node of session l,
i.e., i = s(l), then
fij (l) = r(l). (10.4)
j Ti
If node i is an intermediate relay node for session l, i.e., i = s(l) and i = d(l),
then
fij (l) = fpi (l). (10.5)
j Ti ,j =s(l) pTi ,p=d(l)
It can be easily verified that if (10.4) and (10.5) are satisfied, then (10.6) must
be satisfied. As a result, it is sufficient to list only (10.4) and (10.5) in the
formulation.
Link capacity constraint In addition to the above flow balance equations
at each node i for each session l, the aggregate flow rates on each radio link
cannot exceed this links capacity. To model this mathematically, we need to
first find the capacity on link i j and subband (m, k). If node i sends data
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
to node j on subband (m, k), i.e., xij(m,k) = 1, then the capacity on link i j
and subband (m, k) is given by
% &
(m,k) gij
cij = u(m,k) W (m) log2 1 + ,
where is the ambient Gaussian noise density. Note that the denominator
inside the log function contains only . This is due to one of our interference
constraints stated earlier, i.e., when node i transmits to node j on subband
(m, k), then all the other neighbors of node j within its interference range are
prohibited from using this sub-band. This interference constraint significantly
(m,k) (m,k)
helps simplify the calculation of the link capacity cij . When xij = 0, we
(m,k) (m,k)
have cij = 0. Thus, cij can be written in the following compact form:
% &
(m,k) (m,k) gij
cij = xij u (m,k)
W (m)
log2 1 + . (10.7)
Now, returning to our earlier requirement that the aggregate data rates on
each link i j cannot exceed the links capacity, we have
(m)
K (m,k)
fij (l) cij
lL,s(l)=j,d(l)=i mMij k=1
K
(m) % &
gij
= xij(m,k) u(m,k) W (m) log2 1 + .
mMij k=1
xij(m,k) + (m,k)
xpq 1(i N , m Mi , j Tim ,
qTpm
(i N , j Ti )
fij (l) = r(l) (l L, i = s(l))
j Ti
j =
s(l) p=
d(l)
fij (l) fpi (l) =0 (l L, i N , i = s(l), d(l))
j Ti pTi
(m,k)
xij = 0 or 1, u(m,k) 0 (i N , m Mi , j Tim , 1 k K (m) )
fij (l) 0 (l L, i N , i = d(l), j Ti , j = s(l)),
(m,k)
where W (m) , gij , , , and r(l) are all constants, and xij , u(m,k) , and fij (l)
are the optimization decision variables.
The above optimization problem is a mixed-integer nonlinear programming
(MINLP) problem, which is NP-hard in general [46]. Our approach to solve
this problem is as follows. In Section 10.4, we first propose the computation of
a lower bound for the problem, which can be obtained by relaxing the integer
variables and using a linearization technique. Using this lower bound as a per-
formance benchmark, we develop in Section 10.5 a highly effective algorithm
based on the SF procedure that we discussed in Section 10.1. Using exten-
sive simulation results, we show that the SF algorithm can offer solutions with
objective values very close to the computed lower bounds. Since the optimal
objective value lies between the lower bound and the solution obtained by the
SF algorithm, the solution produced by the SF algorithm must be even closer
to the optimum.
The complexity of the problem formulated in Section 10.3.2 arises from the
(m,k) (m,k)
binary xij -variables and the product of variables xij u(m,k) . To derive
a lower bound for the problem, we first multiply (10.8) and (10.9) by the
(m,k)
corresponding u(m,k) , so that xij appears throughout as a product with
(m,k)
u(m,k) . We then relax the integer (binary) requirement on xij with 0
xij(m,k) 1 and replace xij(m,k) u(m,k) with a single variable, say sij(m,k) , i.e.,
(m,k)
sij = xij(m,k) u(m,k) u(m,k) . Such a relaxation leads to the following lower-
bounding problem formulation:
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
(m)
K
Mininize W (m) sij(m,k)
iN mMi j Tim k=1
(m)
K
subject to u(m,k) = 1 (m M)
k=1
(m,k)
siq u(m,k) 0 (i N , m Mi , 1 k K (m) ) (10.10)
qTi m
(m,k)
sij + (m,k)
spq u(m,k) 0 i N , m Mi , j Tim ,
qTpm
1 k K (m) , p Ijm , p = i (10.11)
lL K
(m) % &
gij (m,k)
fij (l) W (m)
log2 1+ sij 0
s(l)=j,d(l)=i mMij k=1
(i N , j Ti )
fij (l) = r(l) (l L, i = s(l))
j Ti
j =
s(l) p=
d(l)
fij (l) fpi (l) = 0 (l L, i N , i = s(l), d(l))
j Ti pTi
(m,k)
u(m,k) , sij 0 (i N , m Mi , j Tim , 1 k K (m) )
1. Fix the binary values for the x-variables iteratively through a sequence of
LPs.
2. Once all the x-variables are fixed, find a solution (to determine how to
divide sub-bands and the flow routing) based on the x-variable values
obtained in Step 1.
Now, having fixed some x-variables in the first iteration, we update the prob-
lem to obtain a new LP problem for the second iteration as follows. For those
(m,k) (m,k)
xij -variables that are already fixed to one, since sij = xij(m,k) u(m,k) =
u(m,k) , we can replace the corresponding sij(m,k) by u(m,k) . For those xiq
(m,k)
and
(m,k) (m,k) (m,k)
xpq that are fixed to zero, we can set siq = 0 and spq = 0. As a result,
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
1.4
Normalized objective
(w.r.t. lower bound)
1.2
0.8
0.6
0.4
0.2
0
1 10 20 30 40 50 60 70 80 90 100
Set index
all the terms in the LP problem involving these s-variables can be removed and
the corresponding constraints in (10.10) and (10.11) can also be removed.
In the second iteration, we solve this new LP problem and then fix some
additional x-variables based on the same process (now the ordering of the
x-values is done only for the remaining not-yet-fixed or free x-variables).
The iteration continues and eventually we fix all the x-variables to either zero
and one.
Upon fixing all the x-variable values, the original MINLP reduces to an LP
problem, which can be solved in polynomial time. The complete SF algorithm
is displayed in Fig. 10.3.
required that at most one binary x-variable equals one, whereas in the relaxed
(m,k)
problem, there is at most one fraction siq /u(m,k) that exceeds 0.5. Thus,
> 0.5 ensures that both the constraints (10.8) and (10.9) (interference con-
straints at each node and among the nodes) will hold during the SF procedure.
We used = 0.85 in our numerical results in the next section. In the case that
none of the x-variables exceeds , we will fall back to the basic algorithm in
Fig. 10.3 and simply choose the largest valued x-variable.
In this section, we present numerical results for our SF algorithm and com-
pare it to the lower bound that we obtained in Section 10.4. We consider
|N | = 20, 30, or 40 nodes in a 500 500 area (in meters). Among these nodes,
there are |L| = 5 active sessions, each having a rate that is randomly generated
within [10, 100] Mb/s.
We assume that there are M = 5 bands that can be used for the entire net-
work (see Table 10.2). Bands I and II are among the least-utilized spectrum
bands found in [106] (less than 2%), and bands III, IV, and V are unlicensed
ISM bands used for 802.11. Recall that the set of available bands at each CR
node is a subset of these five bands based on the nodes location, and that the
set of available bands at any two nodes in the network may not be identical. In
the simulation, this is done by randomly selecting a subset of bands from the
pool of five bands for each node. Further, we assume that bands I to V can be
divided into 3, 5, 2, 4, and 4 subbands, although other desirable divisions can
be used. Note that the size of each subband may be unequal and is part of the
optimization problem.
We assume that the transmission range at each node is 100 m and that the
interference range is 150 m, although other settings can be used. The path-loss
index is assumed to be 4 andwe set = 62.5. The threshold T is assumed
to be 10. Thus, we have I = 150 T , and the transmission power spectral
100
Note that it is possible that there is no feasible solution for a specific data
set. This could be attributed to the loss of connectivity in certain areas in the
network (due to random network topology), resource bottleneck in a hot area,
etc. Thus, we only report results based on those data sets that have feasible
solutions.
We first present simulation results for 100 data sets for 20-node networks
that can produce feasible solutions. For each data set, the network topology,
the sourcedestination pairs and bit rates of all sessions, and the available fre-
quency bands at each node are randomly generated. We used the SF algorithm
to determine a feasible solution, and we compared its objective value with the
lower bound derived as in Section 10.4. The run time consumed by each sim-
ulation was less than 10 seconds on a Pentium 3.4 GHz machine.
Fig. 10.3 depicts the normalized objective values obtained by the SF algo-
rithm with respect to the computed lower bounds for 100 data sets. The average
normalized objective value obtained among the 100 simulations is 1.04 and the
standard derivation is 0.07. There are two observations that can be made from
this figure. First, since the ratio of the solution value obtained by SF (upper
bound on the optimal solution value) to the lower bound value is close to 1
(in many cases, they coincide with each other), the lower bound must be very
tight. Second, since the actual (unknown) optimal solution value lies between
the solution value obtained by the SF algorithm and the lower bound, the SF
solution value must be even closer to the optimal value than that indicated by
the foregoing ratio.
To get a sense of how the actual (rather than the normalized) numerical
results appear in the simulations, we list the 100 sets of results in Table 10.3.
Note that in many cases, the result obtained by the SF algorithm is identical
to the respective lower bound obtained via the underlying relaxation, which
indicates that the solution found by SF in such cases is optimal.
Simulation results for the 100 random data sets corresponding to the 30-
node and the 40-node networks that produce feasible solutions are shown in
Figs. 10.4 and 10.5, respectively. For the 30-node networks, the average nor-
malized value obtained was 1.10, with a standard derivation of 0.16. For the
40-node networks, the average normalized value obtained was 1.18, with a
standard derivation of 0.16. Thus, the derived SF solutions are again close to
optimality for these data sets.
Table 10.3 Simulation results (in MHz) of 100 data sets for 20-node networks.
1.4
Normalized objective
(w.r.t. Lower bound)
1.2
0.8
0.6
0.4
0.2
0
1 10 20 30 40 50 60 70 80 90 100
Set index
1.4
Normalized objective
(w.r.t. lower bound)
1.2
0.8
0.6
0.4
0.2
0
1 10 20 30 40 50 60 70 80 90 100
Set index
solutions produced by this SF algorithm can offer objective values that are very
close to the computed lower bounds, thus confirming their near-optimality.
10.8 Problems
10.1 For the problem considered in this chapter, what is the purpose of
developing a lower bound? If a solution obtained by the SF algorithm is very
close to the lower bound, what can we claim?
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CHAPTER
11 Metaheuristic methods
One day your life will flash before your eyes. Make sure it is worth
watching.
Unknown
262
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Evaluation
Meet Yes
termination
criteria
No
Selection Stop
q (1q)
Crossover
m
(1m)
Mutation
Another well-known technique is tabu search (TS), which explicitly uses the
history of the search (recorded in a list of solutions declared as tabu) both
to escape from local optima and to implement an exploratory strategy [51].
Similar to SA and iterative improvement, the TS approach is a single-point-
search metaheuristic (also called trajectory methods). The performance of
TS strongly depends on the initial solution and the neighborhood structure,
and on the diversification and intensification schemes that it uses to explore
new or previous promising portions of the search region, respectively.
Unlike the above three metaheuristic methods, a genetic algorithm (GA) is a
population-based algorithm that is inspired by the survival-of-the-fittest prin-
ciple , as derived from its natural evolution context. Because of this algorithmic
structure, a GA is also classified as an evolutionary algorithm. This procedure
has the intrinsic strength of dealing with a set of solutions (i.e., a population) at
each generation rather than working with a single, current solution. To repre-
sent a solution in a GA, we need to first define a data structure (i.e., a gene) and
then map a solution to a sequence of genes (i.e., a chromosome). A GA needs a
set of solutions (or individuals) as the first generation elements. This is the ini-
tial step in Fig. 11.1. A GA evaluates each individual by a fitness function (in
the form of an objective function). If the set of current individuals meets some
termination criteria, then the procedure stops and returns the best individual
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
streams can be used to reconstruct the original video. Further, the quality of
received video commensurate with the number of received descriptions. Such
flexibility has positioned MD coding as a perfect candidate for video commu-
nications over wireless ad hoc networks [102]. This is because the topology of
such networks is intrinsically mesh, within which multiple paths exist between
any source and destination pair. Although the links in such networks are fragile,
the probability of concurrent loss of all of the descriptions is low. Therefore,
MD coding is likely to be very effective.
In this chapter, we study the problem of multi-path routing for double
description (DD) video in wireless ad hoc networks. We follow a cross-layer
approach in problem formulation by considering the application layer perfor-
mance (i.e., average video distortion) as a function of lower layer parameters
(e.g., bandwidth, loss, and path correlation). We show that the objective func-
tion is a complex ratio of high-order exponentials that is not decomposable.
Consequently, it is futile to develop a tractable analytic solution.
However, we find that a metaheuristic technique such as a GA is eminently
suitable in addressing such type of complex cross-layer optimization problems.
This is because GAs possess an intrinsic capability of handling a population of
solutions (rather than working with a single current solution during each itera-
tion). Such capability gives GAs the unique strength of identifying promising
regions in the search space (not necessarily convex) and having less of a ten-
dency to be trapped in a local optimum, as compared with other trajectory-
based metaheuristics (e.g., simulated annealing (SA) and tabu search (TS)
[19]). Using numerical examples, we show that significant performance gains
can be achieved by a GA-based approach over trajectory-based approaches.
The remainder of this chapter is organized as follows. In Section 11.3, we
formulate a cross-layer optimization problem for DD video over multiple paths
in ad hoc networks. In Section 11.4, we describe our GA-based solution. In
Section 11.5, we present numerical examples. Section 11.6 summarizes this
chapter.
Symbol Definition
From these link-level statistics, we can explore path-level bandwidth and fail-
ure probabilities, which are the key factors to determine video distortion (see
(11.2)). Other link characteristics, such as delay, jitter, congestion, and signal
strength may also be incorporated into this framework as well (e.g., see [103]).
Table 11.1 lists the notation used in this chapter.
For video coding and communications, a distortion rate model describes the
relationship between the achieved distortion and the bit rate (i.e., the quality
and the length of the representation). For two descriptions (each generated for
a sequence of video frames), denote dh as the achieved distortion when only
description h is received, h = 1, 2, and d0 the distortion when both descrip-
tions are received. Denote Rh as the rate in bits/pixel of description h, h = 1, 2.
The rate-distortion region for a memoryless independent and identically dis-
tributed (i.i.d.) Gaussian source with the square error distortion measure was
first introduced in [117]. For computational efficiency, Alasti et al. in [6] intro-
duce the following distortion-rate function, which we use in this chapter:
22(R1 +R2 )
d0 = 22R1 +22R2 22(R1 +R2 )
2
d1 = 22R 1 2 (11.1)
2R
d2 = 2 2 2,
where 2 is the variance of the source. Denote P00 as the probability of receiv-
ing both descriptions, P01 as the probability of receiving description 1 only,
P10 as the probability of receiving description 2 only, and P11 as the proba-
bility of losing both descriptions. Then, the average distortion of the received
video can be approximated as:
On the joint portion of the paths, loss of the two streams is correlated.
To model such correlation, we assume that each shared link {i, j } follows
an onoff process modulated by a discrete-time Markov chain, as shown in
Fig. 11.2(a). With this model, there is no packet loss when the link is in
the up state; all packets are dropped when the link is in the down state.
Transition probabilities, {ij , ij }, can be computed from the link statistics as
ij = 1/ lij and ij = (1 pij )/(pij lij ).
For K shared links, the aggregate failure process of these links is a Markov
process with 2K states. In order to simplify the computation, we follow the
well-known Fritchman model [43] to model the aggregate process as an onoff
process. Since a packet is successfully delivered on the joint portion if and only
if all joint links are in the up state, we can lump up all the states with at least
one link failure into a single down state, while using the remaining state
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
bij
(a) The Gilbert two-state link model.
J(P1)
(1)
1 p1dj
S UP DOWN t
(1b)
1 p2dj
J(P1,P2)
J(P2)
(b) A simplified path model for double-description video.
where all the links are in good condition as the up state. Denote Ton as the
average length of the up period. We have
1
Ton = . (11.6)
1 {i, j } J (P1 , P2 ) (1 ij )
If the joint link set is not empty, the probability of a successful delivery on the
joint links can be written as
'
{i, j } J (P1 , P2 ) pij , if J (P1 , P2 ) = ,
pj nt = (11.7)
1 otherwise.
1 pj nt
= and = . (11.8)
Ton Ton (1 pj nt )
With the consolidated path model, the joint probabilities of receiving the
descriptions are
P00 = pj nt (1 ) pdj
1 p2
! dj "
P01 = pj nt pdj
1 1 (1 ) p 2
! " dj
(11.9)
P10 = pj nt 1 (1 )pdj 1 p2
! dj "
P11 = 1 pj nt p 1 + p2 (1 ) p 1 p 2 .
dj dj dj dj
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
4 8
(a) Example ad hoc network.
An individual
P1 1 3 5 7 9
P2 1 2 5 7 8 9
For Problem OPT-MM, each feasible solution consists of a pair of paths (i.e., a
pair of chromosomes), denoted as [P1 , P2 ]. An individual in this case is a pair
of vectors containing the nodes on paths P1 and P2 (see, e.g., Fig. 11.3(b)).
Before entering the main loop in Fig. 11.1, we need to generate an initial
population, i.e., a set of solutions. A simple approach would be to generate this
set of solutions by randomly appending feasible elements (i.e., nodes with con-
nectivity) to a partial solution. Under this approach, each construction process
starts with source node s. Then, the process randomly chooses a link incident
to the current end-node of the partial path and appends the link with its corre-
sponding head-node to augment the path, until destination node t is reached. It
is important to ensure that the intermediate partial path is loop-free during the
process. After generating a certain set of paths for st independently, a pop-
ulation of individuals can be constructed by pairing paths from this set. Our
numerical results show that a properly designed GA is not very sensitive to the
quality of the individuals in the initial population.
11.4.2 Evaluation
The fitness function f (x) of an individual, x = [P1 , P2 ], is closely tied to the
objective function (i.e., distortion D). Since the objective is to minimize the
average distortion function D, we use a fitness function which is defined as
the inverse of the distortion value, i.e., f (x) = 1/D(x). This simple fitness
definition appears to work very well, although other better fitness definitions
may be available.
11.4.3 Selection
During this operation, GA selects individuals that have a better chance or
potential to produce good offspring in terms of their fitness values. By
virtue of the selection operation, good genes among the population are more
likely to be passed to the future generations. We use the so-called Tournament
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Offspring 1
Parent 1
P1 1 3 5 7 9 P1 1 3 5 7 9
crossover operation.
P2 1 2 5 7 8 9 P23 1 2 5 8 9
Parent 2
Offspring 2
P3 1 4 5 8 9 P32 1 4 5 7 8 9
P4 1 3 4 5 6 9 P4 1 3 4 5 6 9
selection [9] scheme, which randomly chooses m individuals from the popula-
tion each time, and then selects the best of these m individuals in terms of their
fitness values. By repeating either procedure multiple times, a new population
can be selected.
11.4.4 Crossover
Crossover mimics the genetic mechanism of reproduction in the natural world,
in which genes from parents are recombined and passed to offspring. The deci-
sion of whether or not to perform a crossover operation is determined by the
crossover rate .
Fig. 11.4 illustrates one possible crossover implementation. Suppose that
we have two-parent individuals x1 = [P1 , P2 ] and x2 = [P3 , P4 ]. We could
randomly pick one path in x1 and one in x2 , say P2 and P3 . If one or more
common nodes exist in these two chosen paths, we could select the first such
common node that exists in P2 , say gr , where gr / {s, t}, and we can then
concatenate nodes {s, . . . , gr } from P2 with nodes {gr+1 , . . . , t} in P3 (where
gr+1 denotes the next downstream node of gr in P3 ) to produce a new path P23 .
Likewise, using the first such node gr in P3 that repeats in P2 (which may be
different from gr ), we can concatenate the nodes {s, . . . , gr } from P3 with
the nodes {gr +1 , . . . , t} in P2 to produce a new path P32 . It is important that
we check the new paths to be sure that they are loop-free. The two offspring
generated in this manner are [P1 , P23 ] and [P32 , P4 ]. On the other hand, if P2
and P3 are disjoint, we could swap P2 with P3 to produce two new offspring
[P1 , P3 ] and [P2 , P4 ].
11.4.5 Mutation
The objective of the mutation operation is to diversify the genes of the current
population, which helps prevent the solution from being trapped in a local
optimum. Just as some malicious mutations could happen in the natural world,
mutation in GA may produce individuals that have worse fitness values. In
such cases, some filtering operation is needed (e.g., the selection operation)
to reject such bad genes and to drive GA toward optimality.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
individual
individual
^
P1 1 3 5 6 9
Original
mutation operation. P1 1 3 5 7 9
New
P2 1 2 5 7 8 9 P2 1 2 5 7 8 9
In this section, we present some numerical examples for the GA-based solu-
tion. In each example, we generate a wireless ad hoc network topology by
placing a number of nodes at random locations in a rectangular region, where
connectivity is determined by the distance coverage of each nodes transmitter.
The sourcedestination nodes s and t are uniformly chosen from the nodes.
For each link, the failure probability is uniformly chosen from [0.01, 0.3]; the
available bandwidth is uniformly chosen from [100, 400] Kb/s, with 50 Kb/s
steps; the mean burst length is uniformly chosen from [2,6]. A DD video codec
is implemented and used in the simulations.
We set the GAs parameters as follows: the population size is 15; = 0.7;
is varied from 0.3 to 0.1 using the schedule described in Section 11.4; 2 is
set to 1, since it does not affect path selection decisions. The GA is terminated
after a predefined number of generations or after a prespecified computation
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
time. The best individual found by the GA upon its termination is prescribed
as the solution to Problem OPT-MM.
11.5.1 Near-optimality
One important performance concern is the quality of the GA solutions. As dis-
cussed, due to the complex nature of Problem OPT-MM, a closed-form optimal
solution is not obtainable. However, for small networks, an optimal solution
may be numerically obtained via an exhaustive search and can be used to com-
pare with the proposed GA-based solutions.
Table 11.2 shows the optimal distortion values found by GA (each is the
average of 30 runs) and by exhaustive search for two ten-node and two 15-
node networks. We find that the solutions found by GA are very close to the
global optimum in all cases. In addition, the standard deviation of the 30 GA
results for the same network is very small. The average computational time
for GA is 0.29 s for the ten-node network (about 60 generations) and 0.39
s for the 15-node network (about 70 generations) on a Pentium 4 2.4 GHz
computer (512 MB memory) with MATLAB 6.5. For exhaustive search, the
average computational time is 58.7 s for the ten-node case and 1877 s for the
15-node case. We also compute the lower bound [104] for each of the networks.
The results are given in the last row of Table 11.2.
Compared with SA, TS explicitly uses the history of the search, both to
escape from local minima and to implement an exploratory strategy. Specif-
ically, TS uses a tabu list to prevent returning to recently visited solutions,
therefore avoiding endless cycling and possibly forcing the search process to
accept nonimproving solutions [51]. In our experiments, we use a tabu list of
five for small networks (e.g., ten-node networks) and ten for large networks
(e.g., 50-node networks). The tabu list is implemented using a first-in-first-out
queue. An explored solution is always inserted at the tail of the queue. When
the queue is full, the head of the queue is removed.
In Fig. 11.6, we plot the evolution of distortion values obtained by GA, SA,
and TS for a ten-node network and a 50-node network, respectively. All the
three metaheuristics are terminated after running for 1 s. Upon termination, GA
has evolved 210 generations in Fig. 11.6(a) and 75 generations in Fig. 11.6(b);
SA ran for 1500 iterations in Fig. 11.6(a) and 700 iterations in Fig. 11.6(b); TS
ran for 1050 iterations in Fig. 11.6(a) and 550 generations in Fig. 11.6(b). GA
has fewer number of iterations than SA and TS. For both networks, the best
distortion values found by GA are evidently much better than those by SA or
TS. In Fig. 11.6(a), GA quickly converges to the global optimal, while both SA
and TS are trapped at local optima (i.e., no further decrease in distortion value
after hundreds of iterations). The same trend can be observed in the 50-node
network case shown in Fig. 11.6(b), although the global optimum cannot be
found here. We also plot the lower bounds in the figures.
An interesting observation from Fig. 11.6 is that for GA, the biggest
improvement in distortion is achieved in the initial iterations, while the
improvement gets smaller as GA evolves more generations. The initial pop-
ulation is generated using the random construction method discussed in Sec-
tion 11.4.1, with no consideration of video performance. The initial solutions
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Average distortion
0.7
0.6
0.5 SA GA TS
0.4
0.3
0.2
0.1 Global optimum Lower bound
0
0 0.2 0.4 0.6 0.8 1
Time (second)
(a) Distortion evolution for a 10-node network.
1
0.9
0.8
SA TS
0.7
Average distortion
0.6
0.5
0.4
0.3
0.2 GA Lower bound
0.1
0
0 0.2 0.4 0.6 0.8 1
Time (second)
(b) Distortion evolution for a 50-node network.
usually have high distortion values. The distortion value quickly drops over
iterations, indicating that the GA performance is not very sensitive to the qual-
ity of the initial population. Also note that the SA and TS curves increase at
some time instances (e.g, the TS curve at 0.06 s in Fig. 11.6(a) and the SA
curve at 0.08 s in Fig. 11.6(b)), which implies that a nonimproving solution is
accepted in order to escape from local minima.
In addition to providing better solutions, another strength of GA over trajec-
tory methods is that multiple good solutions can be found after a single run.
Such extra good paths can be used as alternative (or backup) paths if needed.
to log 1/pij , for all {i, j } E, so that two disjoint paths with the highest
end-to-end success probabilities are used. We compare the performance of our
GA-based multipath routing with these two algorithms over a 50-node ad hoc
network using a video clip.
There are many ways to generate MD video (see [57] for an excellent sur-
vey). We choose a time domain partitioning coding scheme, where two descrip-
tions are generated by separating the even-and odd-numbered frames and cod-
ing them separately. This simple time domain partitioning method is widely
used in many video streaming studies [8; 12; 23; 102]. Compared to a tra-
ditional single description coder, this coder has a comparable computational
complexity. Its coding efficiency is slightly lower than a single description
coder, due to the fact that a longer motion prediction distance is used. However,
this reduced coding efficiency is well justified by the resulting enhanced error
resilience. We use an H.263+ like codec. Since our approach is quite general,
we conjecture that the same trend in performance would be observed for other
video codecs, such as H.264 or MPEG-2 or MPEG-4. The QCIF sequence
Foreman (400 frames) is encoded at 15 fps for each description. A 10%
macroblock level intra-refreshment is used. Each Group of Blocks (GOB) is
carried in a different packet. The received descriptions are decoded and PSNR
values of the reconstructed frames computed. When a GOB is corrupted, the
decoder applies a simple error concealment scheme by copying from the cor-
responding slice in the most recent, correctly received frame.
The quality of the paths found by the three algorithms are presented in
Table 11.3. The 2-SP algorithm has the worst performance in terms of path
success probabilities. The DPSP algorithm has an improved success proba-
bility performance since it uses link success probabilities in routing. How-
ever, it may sacrifice path bandwidth while pursuing low-loss paths. As a
result, it produces the lowest end-to-end bandwidths. We observe that our
GA-based routing yields paths with much higher end-to-end success prob-
abilities and end-to-end bandwidths, resulting in greatly improved video
quality.
The PSNR curves of the received video frames are plotted in Fig. 11.7. We
observe that the PSNR curve obtained by GA is well above those obtained
by the aforementioned network-centric routing approaches. Using GA, the
improvements in average PSNR value over 2-SP and DPSP are 6.29 dB and
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
PSNR (dB)
35
30
25
20
2SP
15
10
5
0
0 50 100 150 200 250 300 350 400
Frame number
(a) GA-based algorithm versus 2-SP.
50
45
40 GA
35
PSNR (dB)
30
25
20
DPSP
15
10
5
0
0 50 100 150 200 250 300 350 400
Frame number
(b) GA-based algorithm versus DPSP.
11.7 Problems
11.1 Describe (in greater detail than in Section 11.1) the basic concepts and
approach of the following metaheuristic methods: iterative improvement, sim-
ulated annealing, tabu search, and genetic algorithm. What is the difference
between the genetic algorithm and these other metaheuristic methods?
11.2 Study the so-called Ant Colony Optimization metaheuristics in the
literature, and compare this with the GA approach.
11.3 Why is the set of initial solutions important to the performance of GA?
11.4 Why is the population size important in GA? What is the consequence
if the population size is too large or too small?
11.5 What is the purpose of crossover in a GA? Explain the impact of the
crossover rate on GA.
11.6 What is the purpose of mutation in a GA? Explain the impact of the
mutation rate on GA.
11.7 We use a dynamic k (see Eq. (11.16)) for mutation. What is the advan-
tage of using such a dynamic k ?
11.8 Why is GA eminently suitable for solving a problem such as Problem
OPT-MM?
11.9 In a GA-based algorithm described in this chapter, how can we ensure
that the solutions after crossover and mutation are feasible?
11.10 What are the different randomization steps within GA? Is randomness
important to the genetic algorithm? Why?
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
PART
IV Other Topics
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
CHAPTER
Luck is dividend of sweat. The more you sweat, the luckier you get.
Unknown
f (n) = (g(n)) if f (n) C g(n) for all n > n0 , where C and n0 are pos-
itive constants;
f (n) = O(g(n)) if f (n) C g(n) for all n > n0 , where C and n0 are pos-
itive constants;
f (n) = (g(n)) if C1 g(n) f (n) C2 g(n) for all n > n0 , where C1 ,
C2 , and n0 are positive constants.
283
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Since the seminal work of Gupta and Kumar [60] on capacity scaling laws
for a wireless network, there has been growing interest in this important area.
Related work on random ad hoc networks can be further divided into the
following two categories: unicast capacity (e.g., [15; 33; 60; 86]) and multi-
cast/broadcast capacity (e.g., [61; 62; 63; 91; 92; 138]). In [60], Gupta and
Kumar showed that for a random network, its capacity is ( B ) under
n ln n
the protocol model, and is [( B ), O( Bn )] under the physical model with
n ln n
synchronized power control, where B is the achievable bit rate for a successful
transmission. In [15; 33; 86], the unicast capacity of multi-channel ad hoc net-
works was investigated under the protocol model. Bhandari and Vaidya [15]
studied the capacity of multi-channel single-radio (MC-SR) networks where
there is a set of channels in the network and each node can only switch to a
subset of these channels. Kyasanur and Vaidya [86] studied the unicast capacity
of multi-channel multi-radio (MC-MR) networks where the number of bands
used at a node is limited by the number of radios at the node. Dai et al. [33]
extended the work of [86] to MC-MR networks with consideration of direc-
tional antennas. Keshavarz-Haddad et al. [61] studied the broadcast capacity
for a homogeneous dense network under the protocol model. The broadcast
capacity under the physical model was studied in [62; 91]. Li et al. [92] exam-
ined the multicast capacity of wireless ad hoc networks. Shakkottai et al. [138]
investigated the multicast capacity of large-scale wireless ad hoc networks
under the protocol model. The multicast capacity under the physical model
was analyzed in [63]. We summarize these efforts in Table 12.1.
Finally, there exists some research related to capacity scaling laws under
specialized wireless communication technologies, such as the generalized
physical model [42; 181] (via adaptive modulation and coding scheme), ultra-
wide band (UWB) [177], multiple-input multiple-output (MIMO) [118], net-
work coding [97], multiple-packet reception (MPR) [165], and cognitive radio
networks [75; 162; 176].
We can see that the protocol model is the simplest model, while the gener-
alized physical model is the most challenging one. In the rest of this chapter,
we study capacity scaling laws under these three interference models. For each
model, we will develop a feasible solution, which will be used as a capacity
lower bound. We will also analyze a capacity upper bound. The results for
these three models are summarized as follows:
Case 1: Under the protocol model, the capacity of a wireless ad hoc network
with n nodes is ( B ) almost surely as n . By almost surely,
n ln n
we mean that the probability of the event that the capacity is ( B )
n ln n
approaches one as n .
1 In [60], to avoid the boundary effect, it is assumed that nodes are deployed on the surface of a
three-dimensional sphere. As we shall see in this chapter, such an assumption is not needed.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Case 2: Under the physical model, the capacity of a wireless ad hoc net-
work with n nodes is (n) [( B ), O( n1/B
)] almost surely as n ,
n ln n
where is the path loss index.
Case 3: Under the generalized physical model, the capacity of a wireless ad
hoc network with n nodes is ( Bn ) almost surely as n .
In Sections 12.3, 12.4, and 12.5, we give details on how to develop these
results. Section 12.6 summarizes this chapter. Table 12.2 lists the notation used
in this chapter.
General notation
D The mean distance between a source node and its destination node
r(n) The common transmission range for all nodes
A parameter to set the interference range
Physical model-specific notation
d(q, b) The length of the q-th hop traversed by the bth unit of data
gij The channel gain on link (i, j )
h(b) The number of hops traversed by the b-th unit of data from its source
to the destination
H = nT b=1 h(b), i.e, the total number of hops traversed by all data in
a time duration T
pij (t) The transmission power used by node i to transmit to node j at time t
under an independent power control
An SINR threshold for successful transmissions
Generalized physical model-specific notation
In this section, we analyze the capacity scaling law under the protocol model.
Under this model, a node i can successfully transmit data with a rate B to a
node j if and only if the following constraints hold:
dij r(n),
where dij is the distance between nodes i and j and r(n) is the transmis-
sion range. The above constraint implicitly sets a constraint on next-hop
routing, i.e., it defines a set of possible candidate nodes (within the trans-
mission range) as the next-hop node.
For any other link (k, l) that is active at the same time, it is necessary that
dkj (1 + )r(n), where (1 + )r(n) represents an interference range.
This is to keep the concurrently transmitting node k sufficiently far away
from node j so that ks interference on j is negligible. More formally, we
have
dkj (1 + )r(n) for each link (k, l) E(t) and (k, l) = (i, j ), (12.1)
where E(t) is the set of links in the network that are active at time t. Con-
straint (12.1) implicitly sets a constraint on scheduling, i.e., it defines conflict
relationships among all the links in the network.
Note that under the protocol model, we assume that the same (synchronized)
transmission power p(n) is used for all nodes. Thus, the transmission range
r(n) is the same for all nodes.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
We consider a common throughput (n) for each node (source) in the net-
work to its randomly selected destination. The goal is to find the maximum
(n) that can be transported by the network. In Section 12.3.1, we develop a
capacity upper bound. The lower bound analysis is given in Section 12.3.2.
The results of this section can be summarized as follows (Theorems 12.1 and
12.2):
Under the protocol model, the capacity of a wireless ad hoc network with n
nodes is (n) = ( B ) almost surely as n .
n ln n
Theorem 12.1
Under the protocol model, a capacity upper bound is O( B ) almost
n ln n
surely as n .
almost surely as n .
Theorem 12.2
Under the protocol model, we can construct a feasible solution with a
throughput of (n) = ( B ) almost surely as n .
n ln n
We now give details for the routing and scheduling schemes described in the
above discussion.
0 0
area of a cell is a(n) = lnnn lnnn = lnnn .2 We set the transmission range
r(n) = 5a(n) so that a node in one cell can transmit to a node in any of its
four neighboring cells. Next, we draw a line to connect each sourcedestination
pair, which passes through some cells. One node is chosen from each of these
cells to relay the traffic from the source node to its destination (see Fig. 12.1
as an example).
Such a routing scheme requires at least one node in each cell. We call a cell
without any node an empty cell. We have the following result regarding the
nonexistence of an empty cell as n :
Lemma 12.1
If the cell size is a(n) = lnnn , an empty cell is nonexistent almost surely as
n . Therefore, the routing scheme is feasible almost surely as n .
2 To divide the unit square into an integral number of small cells, we need to set
a(n) = 1/n/ ln n. However, to make the proofs in this section easy to understand, we
neglect such integrality requirements and simply let a(n) = 1/(n/ ln n) = ln n/n. Note that all
the results in this section hold true if we set a(n) = 1/n/ ln n. For example, we ask the
readers to prove in Exercise 6.10 that Lemma 12.1 still holds when a(n) = 1/n/ ln n.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
D B
Xi
% & % & n
n ln n n n ln n ln n ( ln n)
lim 1 = lim 1
n ln n n n ln n n
n ln n
= lim e
n ln n
1
= lim
n ln n
= 0,
For this routing scheme, we obtain the following result for the number of
S-D lines passing through any cell:
Lemma 12.2
The number of S-D lines passing through any cell is O(n a(n)) almost
surely as n .
Lemma 12.3
For any 10 i n and 1 j n, there exists a constant c1 > 0 such that
pij c1 ln n
n .
0
Proof. Note that cell Qj is contained in a disk of radius dr = 1 l(n) = ln2nn
2
that is centered at Qj s center D (see Fig. 12.2). Denote Xi and Yi as the
source and destination nodes of the S-D pair i, respectively. Note that Xi can
fall either inside or outside the disk.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
We first consider the scenario when Xi falls outside the disk. Suppose that
Xi is at distance x from the disk. We extend the two tangent
lines origination
from Xi equally such that |Xi A| = |Xi B| and Xi C = 2, where C is the mid-
point of line AB. Then Li passes through Qj only if Yi is in the shaded area.
Its area is less than the minimum of 1 (the area of the unit square) and the area
of the triangle, which is equal to 2 2 2 < 2dx r .
(x+dr ) dr
Since Xi is uniformly distributed, the probability density that it is at a dis-
tance x away from the disk is upper bounded by c2 (x + dr ) for some constant
c2 > 0. Thus, we have
Denote Zj = ni=1 Yij as the total number of S-D lines passing through cell
Qj . Consider a fixed j , i.e., a particular cell Qj . For any s, a > 0, based on
the Chernoff bound, we have
E[eaYj ]
P r{Zj > s} . (12.4)
eas
For E[eaZj ], we have
n >
n >
n
E[eaZj ] = E[ea i=1 Yij ]= E[eaYij ] = (ea pij + e0 (1 pij ))
i=1 i=1
>
n
= (1 + (ea 1)pij )
i=1
- .n
=
ln n
1 + (e 1)c4
a
n
- # = $.n
ln n
exp (ea 1)c4
n
= exp(c4 (ea 1) n ln n),
now establish the following lemma on the number of conflicting links for any
link in the network:
Lemma 12.4
Under the cell-based routing scheme, the number of conflicting links for
any link is upper bounded by O(n a(n)) almost surely as n .
Proof. Note that for a link (i, j ), there are two cases in which a link (k, l) could
be its conflicting link. One case is that link (k, l) interferes with link (i, j ) and
the other case is that link (k, l) is interfered by link (i, j ). In this proof, for a
given link (i, j ), we first analyze the number of conflicting links that interfere
with link (i, j ). Since the routing scheme is cell-based, the conflict analysis is
also based on cells. That is, we first analyze the number of interfering cells and
then analyze the number of conflicting links in an interfering cell. We say that
a cell Q is an interfering cell with respect to cell Q if and only if the distance
between a transmitting node in cell Q and a receiving node in cell Q is no
more than (1 + )r(n).
We now show that the number of interfering cells with respect to cell Q
is a constant. For a receiving node j in cell Q, the transmitting nodes of the
links that interfere with j must be within the area inside the solid line shown in
Fig. 12.3, which consists of five rectangles (including the small cell a(n)
a(n) at the center) and four quarter disks with radius (1 + )r(n). To obtain
an upper bound, we define the outermost square area as the interfering area that
will not have any transmitting node in it. Hence, the number of interfering cells
is no more than c6 = (2 (1+)r(n)
a(n)
+ 1)2 = (2 5(1 + ) + 1)2 , which is
a constant.
Next, we analyze the number of links that interfere with link (i, j ) in
each interfering cell. Clearly, this number is no more than the number of
transmissions in this cell. Based on Lemma 12.2 and the routing scheme
discussed earlier, the number of transmissions in a cell is equal to the number
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
of sourcedestination lines intersecting this cell, which is O(n a(n)) almost
surely as n .
Combining the above results, the number of conflicting links that interfere
with link (i, j ) is given by c6 O(n a(n)) = O(n a(n)) almost surely as
n .
Following a similar analysis, we can obtain the same result on the
number of conflicting links that are interfered by link (i, j ). Then the
number of all conflicting links for a link (i, j ) is upper bounded by
O(n a(n)) + O(n a(n)) = O(n a(n)) almost surely as n .
In this section, we analyze the capacity scaling law under the physical model.
Under this model, each node is allowed to perform power control. Denote
pij (t) as the power used by node i to transmit to node j .3 A transmission
with a data rate B is successful if and only if the SINR satisfies
gij pij (t)
SINRij = (k,l)=(i,j ) , (12.5)
+ (k,l)E(t) gkj pkl (t)
3 In [60], the transmission power at all nodes in a random network is the same and is determined
by the number of nodes. A different capacity upper bound O( 1/2 B ) was obtained for this
n
special case. In this section, we study a general case where independent power control is
allowed at each node.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
where gij = dij is the channel gain over link (i, j ), 2 is the path-loss
index, is the ambient noise power, and is the SINR threshold.
In Sections 12.4.1 and 12.4.2, we develop an upper bound and a lower
bound, respectively. The main result can be summarized as follows (also see
Theorems 12.3 and 12.4):
Under the physical model, the capacity of a wireless ad hoc network with n
nodes is (n) [( B ), O( n1/
B
)] almost surely as n .
n ln n
Theorem 12.3
Under the physical model, the capacity of a wireless ad hoc network with n
nodes is (n) = O( n1/
B
) almost surely as n .
(n)nT h(b)
ACD T = d(q, b) (n)nT D, (12.6)
b=1 q=1
where D is the average distance between the source and destination nodes.
Next, we develop an upper bound for ACD T . Since the function f (x) = x
is convex for any 2, we have
(n)nT h(b)
1 1
(n)nT h(b)
d(q, b) d (q, b),
H H
b=1 q=1 b=1 q=1
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
where H = nT b=1 h(b) is the total number of hops traversed by all the data
over time duration T . This gives us
(n)nT h(b)
ACD T = d(q, b)
b=1 q=1
1
1
(n)nT h(b)
H d (q, b)
H
b=1 q=1
1
(n)nT h(b)
=H 1 1 d (q, b) .
(12.7)
b=1 q=1
h(b)
We now analyze H and (n)nT b=1 q=1 d (q, b). We first obtain an upper
bound for H . Note that due to half-duplex, a node that is receiving cannot
transmit at the same time. Therefore, in any given time slot, at most n2 nodes
can transmit simultaneously. For each link (i, j ), its capacity is B. Thus, the
total bits that can be transmitted by all nodes over T is at most nBT
2 , i.e.,
nBT
H . (12.8)
2
(n)nT h(b)
We next find an upper bound for b=1 q=1 d (q, b). Consider a trans-
mission from node i to node j . Since the SINR constraint (12.5) holds at node
j , we have
pij (t)/dij
(k,l)=(i,j ) .
+ (k,l)E(t) pkl (t)/dkj
pij (t)/dij
.
+
(k,l)E(t) pkl (t)/dkj +1
+1 pij (t)
dij pkl (t)
+ (k,l)E(t)
dkj
+1 pij (t)
p
kl (t)
+ (k,l)E(t)
( 2)
+1 pij (t)
< p
,
kl (t)
(k,l)E(t) ( 2)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
where the second inequality holds because dkj 2. Summing the above
inequality over all simultaneous transmissions, we have
+1 (i,j )E(t) pij (t) +1
dij < p (t)
= ( 2) .
(k,l)E(t)
kl
(i,j )E(t) ( 2)
In addition, we have
(n)nT h(b)
T
T
d (q, b) = (dij B) = B dij , (12.10)
b=1 q=1 t=1 (i,j )E(t) t=1 (i,j )E(t)
where the first equality holds since each active link (i, j ) can carry B bits of
data in a unit time. Combining (12.9) and (12.10), we have
(n)nT h(b)
+1
d (q, b) < BT ( 2) . (12.11)
b=1 q=1
Given that we have developed a feasible solution for the protocol model,
we will try to develop a feasible solution for the physical model. We observe
that if we set the parameter in the protocol model to be large enough, then
is also a feasible solution for the physical model. This is because a large
will impose more constraints within the conflict graph, and thus reduce inter-
ference from neighboring nodes in the protocol model solution. As a result,
the SINR at a receiver can be made large enough to satisfy the physical model
SINR constraint (12.5). In this case, is also a feasible solution for the phys-
ical model. The following theorem is based on this insight:
Theorem 12.4
Under the physical model, we can construct a feasible solution with
(n) = ( B ) almost surely as n .
n ln n
interference at the receiving node j from each of these links will be at most
p(n)
[(1+)r(n)] . Therefore, we have
p(n)
r(n)
SINRij > p(n)
+ 1
2[(1+)r(n)]2 [(1+)r(n)]
r(n)
=
p(n)1 + 12 [(1 + )r(n)](2+)
r(n)
= ,
p(n)1 + [ 1 r(n) p(n)1 ]
where the second inequality holds by (12.13). Thus, for solution , the SINR
constraint (12.5) holds for each link (i, j ) that is active at time t. There-
fore, the power control and scheduling schemes in are also feasible for
the physical model. Furthermore, the achieved capacity on each link under
the physical model is equal to that under the protocol model. As a result,
the link capacity constraint is still satisfied for each link under the physi-
cal model, and the achieved throughput under the physical model will be the
same as that under the protocol model, which is ( B ) almost surely as
n ln n
n .
Theorem 12.5
Under the generalized physical model, the capacity of a wireless ad hoc
network generated by a Poisson point process with density n in a 1 1 area
is ( Wn ) almost surely as n , where W is the channel bandwidth.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
301 12.5 Case 3: Asymptotic capacity lower bound under the generalized physical model
Remark 12.1
The above result was established in [42]. Based on this result, the authors
of [42] claimed that this result closes the capacity gap for random networks
since it matches with the upper bound developed in [60]. Unfortunately, this
claim is incorrect for the following reasons:
4 In previous sections, the number of nodes in the network is n. In this section, the number of
nodes in the network is a random number with an expectation value of n.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
l(n)
si
We now provide a sketch of the results. First, we consider the data rate limit
in Phases (ii) and (iii). The constructed highway system has hops with a con-
stant transmission rate. We will also show that the number of nodes that have
access to a highway path is at most O( n). Thus, each source node has a data
rate no more than (W/ n), where W is the channel bandwidth. Second, we
consider the data rate limit in Phases (i) and (iv). We show that the length of
these single hops is at most O(1/ n), which in turn yields a rate ( (lnWn)3 ).
Combining both results, we establish Theorem 12.5.
5 1
< q < 1 3 when n . (12.14)
6 6e
For the number of nodes in a cell, we have the following lemma:
Lemma 12.5
There are less than ln m(n) nodes in each cell almost surely as n .
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
303 12.5 Case 3: Asymptotic capacity lower bound under the generalized physical model
Lemma 12.6
For a Poisson random variable X with parameter , we have
e (e)x
P r(X x) for x > .
xx
e (e)x
P r(X x) exx ln(x/) =
xx
for x > .
ln m(n)
ln m(n)
n ne3
= e 2(m(n))2
2(m(n))2 ln m(n)
= e(1) (ln n) ln n ,
which goes to 0 as n .
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
R2m
m(n)
Rmk ln m(n)
We draw m(n) horizontal lines and m(n) vertical lines across half of the
cells, as shown in Fig. 12.4(b). A path includes some segments of these lines. If
a path does not cross any empty cell, then we call it an open path. We will show
that there are at least ( n) open paths crossing the network area between left
and right sides and at least ( n) crossing paths between top and bottom sides
almost surely as n . We call these paths the highway system. Along the
paths of the highway system, we select one node per cell that acts as a relay.
We now present a proof for crossing paths between left and right sides. The
proof for crossing paths between top and bottom sides is similar and is thus
omitted. We divide the network area into rectangles Rm i of size 1 ln m(n)
m(n)
(see Fig. 12.5), where the constant is given by = * m(n) +. We
m(n) ln(66q)
ln m(n) 3 ln m(n)
note that
m(n) 3
> = > 1, (12.17)
ln m(n) m(n) ln(66q)
3 ln m(n)
ln(6 6q)
where the last equality holds because q < 1 6e13 by (12.14). Denote Pmi as
the maximal number of edge-disjoint crossing paths
(between left andright
m(n)
sides) within the rectangle Rm , and let Pm = min Pm : 1 i ln
i i
m(n) . We
have the following result:
Theorem 12.6
ln(66q)3
For a constant = q
ln 1q
> 0, we have that
where P rq {A}
0 is the probability of event A under a nonempty probability
q, m(n) = * m(n)
2 ln 6 , and =
n + . That is, there are at least
m(n) ln(66q)
ln m(n) 3 ln m(n)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
305 12.5 Case 3: Asymptotic capacity lower bound under the generalized physical model
Lemma 12.7
} of the event
For a nonempty probability q < 13 , the probability P rq {Bwh
Bwh satisfies
P rq {Bwh } (h + 1)eln(3q)w .
Proof. Denote Ai as the event that there exists a crossing path starting from
the i-th vertex on the left side of box Bwh , where vertices on the left side are
}.
ordered from the bottom to the top. Thus, we have i P rq {Ai } = P rq {Bwh
Since the number of vertices is at most h + 1, there is at least one index i0
such that P rq {Ai0 } h+1
1 }, i.e.,
P rq {Bwh
P rq {Bwh } (h + 1)P rq {Ai0 }.
Denote Ni0 (w) as the number of open paths with w hops starting at the
i0 -th vertex. Since a crossing path starting from the i0 -th vertex to the right
side of Bwh has at least w hops, we have
Note that paths with w hops are open with probability q w and the number
of paths with w hops starting at the i0 -th vertex is at most 3w . We have
almost surely for a large number of open paths. Combining the above inequal-
ities, we have
P rq {Bwh } (h + 1)P rq {Ai0 }
(h + 1)P rq {Ni0 (w) 1}
(h + 1)(3q)w
= (h + 1)eln(3q)w
The above lemma asserts that if the nonempty probability is too small, then
a box Bwh cannot be crossed between its left and right sides. We consider
the dual question of the existence of a crossing path by exploiting the concept
of a dual grid from the Percolation theory. A dual grid Sm for the grid Bm is
constructed by placing a vertex in each cell of Bm , and connecting two neigh-
boring vertices by an dashed line, as shown in Fig. 12.6. For each edge in Sm
that does not cross an open edge of the original gird, we call it an open edge
in Sm . We have the following relationship between the original grid and its
dual: if the original grid has a nonempty probability q, then the dual grid has a
nonempty probability 1 q. For the existence of a crossing path, we have the
following lemma:
Lemma 12.8
As long as q > 23 , we have
P rq {Bwh } 1 (w + 1)eln(33q)h .
Proof. For the box Bwh and the corresponding dual box Swh , the comple-
is event S
ment of Bwh wh that there exists an open crossing path in Swh
between its top and bottom sides [80].
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
307 12.5 Case 3: Asymptotic capacity lower bound under the generalized physical model
Note that the dual lattice has a nonempty probability q = 1 q < 13 . After
rotating the box by 90 , we can apply Lemma 12.7 to the dual lattice and obtain
P r{Swh } (w + 1)eln(3q)h = (w + 1)eln(33q)h .
Based on the above results, we are now ready to give a proof of Theo-
rem 12.6.
Proof of Theorem 12.6. Denote Rm i as the event that there exists a crossing
path within the rectangle Rm between its left and right sides and let q = 2q 1.
i
Thus, we have
where the first inequality holds by (12.18), and the second inequality holds by
using ln m(n) ln m(n) and (12.19).
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Therefore, we have
q
ln m(n) ln + ln(m(n) + 1) + ln(6 6q) ln m(n) < ln(1 e m(n) ),
1q
which is equivalent to (12.22).
By (12.21) and (12.22), we have
m(n) 1
P rq {Pm ln m(n)} < 1 e m(n) ln m(n) = 1 e ln m(n) ,
5 Incidentally, we fix an error in [42] (Eq. (15)), where it was incorrectly stated that
m(n)
i ln m(n)}) ln m(n) .
P rq {Pm ln m(n)} = (P rq {Pm
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
309 12.5 Case 3: Asymptotic capacity lower bound under the generalized physical model
Source
k=2(d+1)
=4
A lower bound for per-node throughput We can derive a lower bound for
per-node throughput by analyzing the achievable per-node throughput by the
above routing and scheduling schemes.
For the achievable rate at each hop, we have the following result:
Theorem 12.7
Denote cij (d) as the data rate at which node i in a cell can transmit to node
j in another cell that is at most d (diagonal) cells away. Then cij (d) is at
least (W ) almost surely as n , where d = 1 for Phases (ii) and (iii)
and d = ln m(n) + 1 for Phases (i) and (iv).
Proof. We focus on a particular cell. To analyze cij (d), we first find an upper
bound for the interference at the receiver j . We consider the transmitters in the
eight closest cells that are located at a Euclidean distance of at least l(n)(d +
1) from the receiver (see Fig. 12.8), along with the 16 next closest cells that
are located at a Euclidean distance of at least l(n)(3d + 3), and so on. By
extending the sum of the interferences to the entire two-dimensional area, we
have the following bound:
8iP 8P i
I (d) = .
(l(n)(2i 1)(d + 1)) (l(n)(d + 1)) (2i 1)
i=1 i=1
(12.26)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
311 12.5 Case 3: Asymptotic capacity lower bound under the generalized physical model
S(d) ((l(n)d) )
= (1).
+ I (d) + ((l(n)d) )
S(d)
As a result, the achievable data rate cij (d) = W log2 1 + +I (d) has a con-
stant lower bound (W ) almost surely as n .
We can now analyze the achievable per-node throughput, which is bounded
by the bottleneck phases among Phases (i)(iv). We will show that the commu-
nication bottleneck resides in the highway Phases (ii) and (iii) with a per-node
throughput of ( 1n ).
For Phase (i), we have the following lemma:
Lemma 12.9
For transmissions
in Phase (i), each source node can achieve at least a rate
(ln n)3 almost surely as n .
W
Proof. To compute the rate at which nodes can communicate to the entry
points, we apply the second part of Theorem 12.7, which says that each node
can communicate to its entry point at least at a rate (W ). For transmissions in
Phase (i), where d = ln m(n) + 1, there are (2(d + 1))2 time slots required
for scheduling. By Lemma 12.5, the number of nodes in each cell is less than
ln m(n) almost surely as n . Therefore, the actual rate available
for each
node is at least (W )/((2(d + 1)) ln m(n)) = d 2 ln m(n) , i.e., the per-
2 W
node throughput is (lnWn)3 almost surely as n .
Since each node is the destination of exactly one source, we can prove the
following lemma by adopting the same approach as that in Lemma 12.9:
Lemma 12.10
For transmissions in Phase (iv), each destination
node can receive data from
the highway at least at a rate (ln n)3 almost surely as n .
W
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Next, we consider the data rate along the highway. We first have the follow-
ing lemma:
Lemma 12.11
There are fewer than 2nh
m(n) nodes on each strip almost surely as n .
Proof. Let An be the event that there is at least one strip with no fewer
2nh
than m(n) nodes. Since the number of nodes in each strip is a Poisson ran-
nh
dom variable with parameter m(n) , by the union bounds and Lemma 12.6,
we have
# $ m(n)
2nh
1 nh
nh
e m(n) 1 m(n)
nh
e 2nh 1 e m(n)
nh
P r{An } e m(n)
m(n)
2nh
= e = ,
h m(n)
h 2 h 4
which goes to 0 as n .
The following lemma gives the achievable per-node throughput along the
highway:
Lemma 12.12
For transmissions in Phases (ii) and (iii), each node along the highway may
forward data for many source nodes. The achievable per-node throughput
for each source node is at least ( Wn ) almost surely as n .
Proof. We prove the result for the horizontal traffic in Phase (ii) (the proof for
Phase (iii) is similar and is thus omitted). For transmissions in Phase (ii), we
have d = 1 and the number of time slots required for scheduling is a constant
(2(d + 1))2 = 16. Further, the number of nodes selected for the highway in
each cell is one. Such a node can relay at most all the traffic generated in the
h
corresponding strip of height m(n) . According to Lemma 12.11, a node on a
horizontal highway may relay traffic for all nodes in this strip, which is at most
2nh
m(n) nodes. By applying the first part of Theorem 12.7, we conclude that one
source nodes data along the highway are at least ( W2nh ), i.e., the per-node
161 m(n)
rate is ( W
2nh ) = ( W m(n)
n ) = ( n ), almost surely as n .
W
m(n)
12.7 Problems
12.1 Analyze the (), O(), and () relationships between the following
f (n) and g(n) functions:
(a) f (n) = 3n2 + 2n, g(n) = n2 ;
(b) f (n) = ln n, g(n) = n;
(c) f (n) = 1/n, g(n) = 1.
12.2 Describe the details for the protocol model, the physical model, and
the generalized physical model that we used in the asymptotic analysis in this
chapter.
12.3 For any two simultaneously active links i j and k l under the
protocol model, what is the maximum ratio between the interference from node
k to node j and the signal from node i to node j ?
12.4 In the proof of Theorem 12.1, we established that the network can
support O(r 2 (n)) transmissions at any time. Why is this not n (the number
of all source nodes)?
12.5 For a set of links whose receivers are inside a square of side-length
2
1/ r(n) , show that these links interfere with each other under the protocol
model.
12.6 If the cell size is changed to a(n) = 1/ lnnn in Lemma 12.1, prove that
there is still no empty cell almost surely as n . (Hint: You can prove this
by following a similar approach to that in the proof of Lemma 12.1.)
12.7 Display a possible route for the case that the S-D line crosses a gird
point in Fig. 12.9.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Link c
Link d
12.8 Under the protocol model, use a conflict graph to represent the interfer-
ence relationships between the links shown in Fig. 12.10, where the solid lines
represent links and the dotted lines represent interference. For example, there
exists a dotted line between the transmitter of link c and the receiver of link a,
which means that the receiver of link a is within the interference range of the
transmitter of link c . What is maximum node degree in the resulting conflict
graph?
12.9 What is the relationship between B and under the physical model?
12.10 Show that a capacity upper bound is O( Bn ) when all nodes use the
same transmit power under the physical model.
12.11 Suppose that all nodes use the same transmit power under the physical
model. Show that for any two successful simultaneous transmissions i j
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
12.14 Compare the upper bound for a general case of independent power
control (in Theorem 12.3) and the upper bound for a special case of synchro-
nized power control (in Footnote 3). Which case has a tighter bound? Why can
we obtain a tighter bound for the other case?
12.15 For a random network generated by a Poisson point process with den-
sity n in a 1 1 area, show that if we divide the 1 1 area into small regions
of area lnnn , then every small region has at least one node almost surely as
n .
12.16 Why do we develop bounds for q in (12.14)? (Hint: Consider where
we have used q.)
12.17 Lemma 12.7 considers the case of q < 13 . Does the same result hold
for the case of q 13 ? If yes, why do we focus on the case of q < 13 in this
lemma? (Hint: Consider where we have used Lemma 12.7.)
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
References
1 E. Aarts and J. Korst, Simulated Annealing and Boltzman Machines, New York, NY,
John Wiley & Sons, 1989.
2 C. Ahn and R. Ramakrishna, A genetic algorithm for shortest path routing problem
and the sizing of populations, IEEE Transactions on Evolutionary Computation, vol. 6,
no. 6, pp. 566579, December 2002.
3 R.K. Ahuja, Algorithms for the minimax transportation problem, Naval Research
Logistics Quarterly, vol. 33, pp. 725739, 1986.
4 I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, Wireless sensor net-
works: a survey, Computer Networks (Elsevier), vol. 38, no. 4, pp. 393422, 2002.
5 I. Akyildiz and X. Wang, A survey on wireless mesh networks, IEEE Communications
Magazine, vol. 43, no. 9, September 2005.
6 M. Alasti, K. Sayrafian-Pour, A. Ephremides, and N. Farvardin, Multiple description
coding in networks with congestion problem, IEEE Transactions on Information The-
ory, vol. 47, no. 3, pp. 891902, March 2001.
7 M. Alicherry, R. Bhatia, and L. Li, Joint channel assignment and
routing for throughput optimization in multi-radio wireless mesh net-
works, in Proc. ACM MobiCom, pp. 5872, Cologne, Germany, August
28September 2, August 2005.
8 J.G. Apostolopoulos, T. Wong, W. Tan, and S. Wee, On multiple description streaming
in content delivery networks, in Proc. IEEE INFOCOM, pp. 17361745, New York,
NY, June 2002.
9 T. Back, D. Fogel, and Z. Michalewicz (eds), Handbook of Evolutionary Computation,
New York, NY, Oxford University Press, 1997.
10 M.S. Bazaraa, J.J. Jarvis, and H.D. Sherali, Linear Programming and Network Flows,
4th edition, New York, NY, John Wiley & Sons Inc., 2010.
11 M.S. Bazaraa, H.D. Sherali, and C.M. Shetty, Nonlinear Programming: Theory and
Algorithms, 3rd edition, New York, John Wiley & Sons, 2006.
12 A.C. Begen, Y. Altunbasak, and O. Ergun, Multi-path selection for multiple descrip-
tion encoded video streaming, EURASIP Signal Processing: Image Communications,
vol. 20, no. 1, pp. 3960, January 2005.
13 A. Behzad and I. Rubin, Impact of power control on the performance of ad hoc wire-
less networks, in Proc. IEEE INFOCOM, pp. 102113, Miami, FL, March 1317,
2005.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
317 References
14 D. Bertsekas and R. Gallager, Data Networks, Englewood Cliffs, NJ, Prentice Hall,
1992.
15 V. Bhandari and N.H. Vaidya, Capacity of multi-channel wireless networks with ran-
dom (c, f ) assignment, in Proc. ACM MobiHoc, pp. 229238, Montreal, QC, Canada,
September 914, 2007.
16 R. Bhatia and M. Kodialam, On power efficient communication over multi-hop wire-
less networks: joint routing, scheduling and power control, in Proc. IEEE INFOCOM,
pp. 14571466, Hong Kong, China, March 711, 2004.
17 E. Biglieri, R. Calderbank, A. Constantinides, A. Goldsmith, A. Paulraj, and H.V.
Poor, MIMO Wireless Communications, Cambridge, Cambridge University Press, Jan-
uary 2007.
18 A. Bletsas, A. Khisti, D. Reed, and A. Lippman, A simple cooperative diversity method
based on network path selection, IEEE Journal on Selected Areas in Communications,
vol. 24, no. 3, pp. 659672, March 2006.
19 C. Blum and A. Roli, Metaheuristics in combinatorial optimization: overview and con-
ceptual comparison, ACM Computing Surveys, vol. 35, no. 3, pp. 268308, Septem-
ber 2003.
20 T.X. Brown, H.N. Gabow, and Q. Zhang, Maximum flow-life curve for a wireless ad
hoc network, in Proc. ACM MobiHoc, pp. 128136, Long Beach, CA, October 45,
2001.
21 N. Bulusu and S. Jha (eds.), Wireless Sensor Networks: A Systems Perspective, Nor-
wood, MA, Artech House, 2005.
22 J. Cai, S. Shen, J.W. Mark, and A.S. Alfa, Semi-distributed user relaying algorithm for
amplify-and-forward wireless relay networks, IEEE Transactions on Wireless Commu-
nication, vol. 7, no. 4, pp. 13481357, April 2008.
23 J. Chakareski, S. Han, and B. Girod, Layered coding vs. multiple descrip-
tions for video streaming over multiple paths, in Proc. ACM Multimedia,
pp. 422431, Berkeley, CA, November 2003.
24 J.H. Chang and L. Tassiulas, Maximum lifetime routing in wireless sensor networks,
IEEE/ACM Transactions on Networking, vol. 12, issue 4, pp. 609619, August 2004.
25 P. Charnsethikul, The constrained minimax linear assignment problem, Optimization
Methods and Software, vol. 14, issues 1 & 2, pp. 3748, 2000.
26 C.C. Chen and D.S. Lee, A joint design of distributed QoS scheduling and power
control for wireless networks, in Proc. IEEE INFOCOM, Barcelona, Catalunya, Spain,
April 2329, 2006.
27 L. Chen, S.H. Low, M. Chiang, and J.C. Doyle, Cross-layer congestion control, rout-
ing and scheduling design in ad hoc wireless networks, in Proc. IEEE INFOCOM,
Barcelona, Catalunya, Spain, April 2329, 2006.
28 Concorde TSP Solver, http://www.tsp.gatech.edu/concorde/.
29 N. Clemens and C. Rose, Intelligent power allocation strategies in an unlicensed spec-
trum, in Proc. IEEE DySPAN, pp. 3742, Baltimore, MD, November 811, 2005.
30 T.J. Cormen, C.E. Leiserson, R.L. Rivest, and C. Stein, Introduction to Algorithms, 2nd
edition, The MIT Press, 2001.
31 IBM ILOG CPLEX Optimizer, http://www-01.ibm.com/software/integration/
optimization/cplex-optimizer/.
32 R.L. Cruz and A.V. Santhanam, Optimal routing, link scheduling and power control in
multi-hop wireless networks, in Proc. IEEE INFOCOM, pp. 702711, San Francisco,
CA, March 30April 3, 2003.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
318 References
33 H. Dai, K. Ng, R.C. Wong, and M. Wu, On the capacity of multi-channel wireless net-
works using directional antennas, in Proc. IEEE INFOCOM, pp. 13011309, Phoenix,
AZ, April 1318, 2008.
34 I. Dietrich and F. Dressler, On the lifetime of wireless sensor networks, ACM Trans-
actions on Sensor Networks, vol. 5, no. 1, pp. 139, February 2009.
35 R. Draves, J. Padhye, and B. Zill, Routing in multi-radio, multi-hop wireless mesh
networks, in Proc. ACM MobiCom, pp. 114128, Philadelphia, PA, September 26
October 1, 2004.
36 J. Dreo, A. Petrowski, P. Siarry, and E. Taillard, Metaheuristics for Hard Optimization:
Methods and Case Studies, New York, Springer-Verlag, 2006.
37 T. Elbatt and A. Ephremides, Joint scheduling and power control for wireless ad-hoc
networks, in Proc. IEEE INFOCOM, pp. 976984, New York, NY, June 2327, 2002.
38 D. Eppstein, Finding the k shortest paths, SIAM Journal on Computing, vol. 28, no. 2,
pp. 652673, August 1999.
39 R. Etkin, A. Parekh, and D. Tse, Spectrum sharing for unlicensed bands, in
Proc. IEEE DySPAN, pp. 251258, Baltimore, MD, November 811, 2005.
40 FCC (Federal Communications Commission), Facilitating opportunities for flexible,
efficient, and reliable spectrum use employing cognitive radio technologies, notice of
proposed rule making and order, FCC 03-322.
41 G. J. Foschini and M. J. Gans, On limits of wireless communications in a fading envi-
ronment when using multiple antennas, Wireless Personal Communications, vol. 6,
pp. 311355, March 1998.
42 M. Franceschetti, O. Dousse, D.N.C. Tse, and P. Thiran, Closing the gap in the capac-
ity of wireless networks via percolation theory, IEEE Transaction on Information The-
ory, vol. 53, no. 3, pp. 10091018, March 2007.
43 D. Fritchman, A binary channel characterization using partitioned markov
chains, IEEE Transactions on Information Theory, vol. 13, no. 2,
pp. 221227, April 1967.
44 A.E. Gamal, J. Mammen, B. Prabhakar, and D. Shah, Throughput-delay trade-off
in wireless networks, in Proc. IEEE INFOCOM, pp. 464475, Hong Kong, China,
March 711, 2004.
45 G. Ganesan and Y.G. Li, Cooperative spectrum sensing in cognitive radio networks,
in Proc. IEEE DySPAN, pp. 137143, Baltimore, MD, November 811, 2005.
46 M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide to the Theory of
NP-Completeness, New York, NY, W.H. Freeman and Company, 1979.
47 A. Gashemi and E. Sousa, Collaborative spectrum sensing for opportunistic access in
fading environments, in Proc. IEEE DySPAN, pp. 131136, Baltimore, MD, Novem-
ber 811, 2005.
48 D. Gesbert, M. Shafi, D. Shiu, P. J. Smith, and A. Naguib, From theory to practice: an
overview of MIMO space-time coded wireless systems, IEEE Journal on Select Areas
in Communications, vol. 21, no. 3, pp. 281302, April 2003.
49 E. Giler, Eric Giler demos wireless electricity, URL:http://www.ted.com/talks/eric
giler demos wireless electricity.html.
50 A. Giridhar and P.R. Kumar, Maximizing the functional lifetime of sensor networks,
in Proc. ACM/IEEE International Symposium on Information Processing in Sensor Net-
works, pp. 512, Los Angeles, CA, April 2005.
51 F. Glover and M. Laguna, Tabu Search. Boston, MA, Kluwer-Academic, 1997.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
319 References
320 References
71 Y.T. Hou, Y. Shi, and H.D. Sherali, Rate allocation and network lifetime problems
for wireless sensor networks, IEEE/ACM Transactions on Networking, vol. 16, no. 2,
pp. 321334, April 2008.
72 Y.-C. Hu and D. Johnson, Design and demonstration of live audio and
video over multihop wireless ad hoc networks, in Proc. IEEE MILCOM, pp.
710, Anaheim, CA, October 2002.
73 J. Huang, R.A. Berry, and M.L. Honig, Spectrum sharing with distributed interference
compensation, in Proc. IEEE DySPAN, pp. 8893, Baltimore, MD, November 811,
2005.
74 O. Ileri, D. Samardzija, T. Sizer, and N.B. Mandayam, Demand responsive pricing
and competitive spectrum allocation via a spectrum server, in Proc. IEEE DySPAN,
pp. 194202, Baltimore, MD, November 811, 2005.
75 S.-W. Jeon, N. Devroye, M. Vu, S.-Y. Chung, and V. Tarokh, Cognitive networks
achieve throughput scaling of a homogeneous network, in Proc. ICST International
Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks
(WiOpt), Seoul, Korea, June 2627, 2009.
76 X. Jiang, J. Polastre, and D. Culler, Perpetual environmentally powered sensor net-
works, in Proc. ACM/IEEE International Symposium on Information Processing in
Sensor Networks, pp. 463468, Los Angeles, CA, April 2005.
77 A. Kansal, J. Hsu, S. Zahedi, and M.B. Srivastava, Power management in energy har-
vesting sensor networks, ACM Transactions Embedded Computing Systems, vol. 6,
no. 4, article 32, September 2007.
78 K. Kar, M. Kodialam, T.V. Lakshman, and L. Tassiulas, Routing for network capac-
ity maximization in energy-constrained ad-hoc networks, in Proc. IEEE INFOCOM,
pp. 673681, March 30April 3, 2003, San Francisco, CA.
79 F. Kelly, A. Maulloo, and D. Tan, Rate control for communication networks: shadow
prices, proportional fairness and stability, in Journal of the Operational Research Soci-
ety, vol. 49, no. 3, pp. 237252, March 1998.
80 H. Kesten, Percolation Theory for Mathematicians, Boston, MA, Birkhauser, 1982.
81 J. Kleinberg and E. Tardos, Algorithm Design, Boston, MA, Addison-Wesley Longman
Publishing, 2005.
82 M. Kodialam and T. Nandagopal, Characterizing the capacity region in multi-radio
multi-channel wireless mesh networks, in Proc. ACM MobiCom, pp. 7387, Cologne,
Germany, August 28September 2, 2005.
83 M. Kodialam and T. Nandagopal, Characterizing achievable rates in multi-hop wire-
less mesh networks with orthogonal channels, IEEE/ACM Transactions on Network-
ing, vol. 13, no. 4, pp. 868880, August 2005.
84 G. Kramer, I. Maric, and R.D. Yates, Cooperative communications, Foundations and
Trends in Networking, Now Publishers, June 2007.
85 A. Kurs, A. Karalis, R. Moffatt, J.D. Joannopoulos, P. Fisher, and M. Soljacic, Wireless
power transfer via strongly coupled magnetic resonances, Science, vol. 317, no. 5834,
pp. 8386, 2007.
86 P. Kyasanur and N.H. Vaidya, Capacity of multi-channel wireless networks: impact
of number of channels and interfaces, in Proc. ACM MobiCom, pp. 4357, Cologne,
Germany, August 28September 2, 2005.
87 J.N. Laneman, D.N.C. Tse, and G.W. Wornell, Cooperative diversity in wireless net-
works: efficient protocols and outage behavior, IEEE Transactions on Information
Theory, vol. 50, no. 12, pp. 30623080, December 2004.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
321 References
88 L. Lasdon, Optimization Theory for Large Scale Systems, New York, NY, Macmillan.,
1970.
89 E. Lawler, Combinatorial Optimization: Networks and Matroids, Mineola, NY, Dover
Publications Inc., 2001.
90 J.-W. Lee, M. Chiang, and R.A. Calderbank, Price-based distributed algorithm for
optimal rate-reliability tradeoff in network utility maximization, IEEE Journal on
Selected Areas in Communications, vol. 24, issue 5, pp. 962976, May 2006.
91 X.-Y. Li, J.-Z. Zhao, Y.-W. Wu, S.-J. Tang, X.-H. Xu, and X.F. Mao, Broadcast capac-
ity for wireless ad hoc networks, in Proc. IEEE MASS, pp. 114123, Atlanta, GA,
September 29October 2, 2008.
92 X.-Y. Li, Multicast capacity of wireless ad hoc networks, IEEE/ACM Transactions
Networking, vol. 17, no. 3, pp. 950961, June 2009.
93 Z. Li, Y. Peng, W. Zhang, and D. Qiao, J-RoC: A joint routing and charging scheme
to prolong sensor network lifetime, in Proc. IEEE ICNP, pp. 373382, Vancouver,
Canada, October 1720, 2011.
94 D. Linden and T.B. Reddy (eds.), Handbook of Batteries, 3rd edition, McGraw-Hill,
2002.
95 LINDO SystemsOptimization Software, http://www.lindo.com/.
96 X. Liu and W. Wang, On the characteristics of spectrum-agile communication net-
works, in Proc. IEEE DySPAN, pp. 214223, Baltimore, MD, November 811, 2005.
97 J. Liu, D. Goeckel, and D. Towsley, Bounds on the gain of network coding and broad-
casting in wireless networks, in Proc. IEEE INFOCOM, pp. 724732, Anchorage, AK,
May 612, 2007.
98 S.H. Low, A duality model of TCP and queue management algorithms, IEEE/ACM
Transactions on Networking, vol. 11, issue 4, pp. 525536, August 2003.
99 H. Luss and D.R. Smith, Resource allocation among competing activities: a lexico-
graphic minimax approach, Operations Research Letters, vol. 5, no. 5, pp. 227231,
November 1986.
100 J. R. Magnus and H. Neudecker, Matrix Differential Calculus with Applications in
Statistics and Economics, New York, NY, Wiley, 1999.
101 N. Malpani and J. Chen, A note on practical construction of maximum bandwidth
paths, Information Processing Letters, vol. 83, pp. 175180, August 2002.
102 S. Mao, S. Lin, S.S. Panwar, Y. Wang, and E. Celebi, Video transport over ad hoc
networks: multistream coding with multipath transport, IEEE Journal on Select Areas
in Communications, vol. 21, no. 10, pp. 17211737, December 2003.
103 S. Mao, S. Kompella, Y.T. Hou, H.D. Sherali, and S.F. Midkiff, Routing for concur-
rent video sessions in ad hoc networks, IEEE Transactions on Vehicular Technology,
vol. 55, issue 1, pp. 317327, January 2006.
104 S. Mao, Y.T. Hou, X. Cheng, H.D. Sherali, S.F. Midkiff, and Y.-Q. Zhang, On routing
for multiple description video over wireless ad hoc networks, IEEE Transactions on
Multimedia, vol. 8, issue 5, pp. 10631074, October 2006.
105 I. Maric and R.D. Yates, Cooperative multihop broadcast for wireless networks,
IEEE Journal on Selected Areas in Communications, vol. 22, issue 6, pp. 10801088,
August 2004.
106 M. McHenry and D. McCloskey, Spectrum occupancy measurements: Republican
National Convention, New York, NY, August 30September 3, 2004, available at
http://www.sharedspectrum.com/wp-content/uploads/4 NSF NYC Report.pdf.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
322 References
323 References
126 A.P. Punnen and K.P.K. Nair, Improved complexity bound for the maximum cardi-
nality bottleneck bipartite matching problem, Elsevier Discrete Applied Mathematics,
vol. 44, pp. 9193, 1994.
127 R. Ramanathan and R. Rosales-Hain, Topology control of multihop wireless networks
using transmit power adjustment, in Proc. IEEE INFOCOM, pp. 404413, 2000, Tel
Aviv, Israel, March 2630.
128 A. Raniwala and T. Chiueh, Architecture and algorithms for an IEEE 802.11-based
multi-channel wireless mesh network, in Proc. IEEE INFOCOM, pp. 22232234,
Miami, FL, March 1317, 2005.
129 T.S. Rappaport, Wireless Communications: Principles and Practice, Upper Saddle
River, NJ, Prentice Hall, 1996.
130 J.H. Reed, Software Radio: A Modern Approach to Radio Engineering, Upper Saddle
River, NJ, Prentice Hall, May 2002.
131 V. Rodoplu and T.H. Meng, Minimum energy mobile wireless networks, IEEE Jour-
nal on Selected Areas in Communications, vol. 17, no. 8, pp. 13331344, August 1999.
132 J. B. Rosen, The gradient projection method for nonlinear programming, part I: linear
constraints, SIAM Journal on Applied Mathematics, vol. 8, pp. 181217, 1960.
133 A. Sankar and Z. Liu, Maximum lifetime routing in wireless ad-hoc networks, in
Proc. IEEE INFOCOM, pp. 10891097, Hong Kong, China, March 711, 2004.
134 Wireless Innovation Forum, http://www.sdrforum.org.
135 E. Setton, Y. Liang, and B. Girod, Adaptive multiple description video streaming over
multiple channels with active probing, in Proc. IEEE ICME, Baltimore, MD, July
2003.
136 A. Sendonaris, E. Erkip, and B. Aazhang, User cooperation diversity part I: system
description, IEEE Transactions on Communications, vol. 51, no. 11, pp. 19271938,
November 2003.
137 A. Sendonaris, E. Erkip, and B. Aazhang, User cooperation diversity part II: imple-
mentation aspects and performance analysis, IEEE Transactions on Communications,
vol. 51, no. 11, pp. 19391948, November 2003.
138 S. Shakkottai, X. Liu, and R. Srikant, The multicast capacity of large multihop wireless
networks, in Proc. ACM MobiHoc, pp. 247255, Montreal, Quebec, Canada, Septem-
ber 914, 2007.
139 S. Sharma, Y. Shi, Y.T. Hou, and S. Kompella, An optimal algorithm for relay node
assignment in cooperative ad hoc networks, IEEE/ACM Transactions on Networking,
vol. 19, issue 3, pp. 879892, June 2011.
140 H.D. Sherali and W.P. Adams, A hierarchy of relaxations between the continuous and
convex hull representations for zero-one programming problems, SIAM Journal on
Discrete Mathematics, vol. 3, no. 3, pp. 411430, 1990.
141 H.D. Sherali and C.H. Tuncbilek, A global optimization algorithm for polynomial pro-
gramming problems using a reformulation-linearization technique, Journal of Global
Optimization, vol. 2, pp. 101112, 1992.
142 H.D. Sherali and W.P. Adams, A hierarchy of relaxations and convex hull characteri-
zations for mixed-integer zero-one programming problems, Discrete Applied Mathe-
matics, vol. 52, pp. 83106, 1994.
143 H.D. Sherali and C.H. Tuncbilek, Comparison of two reformulation-linearization tech-
nique based linear programming relaxations for polynomial programming problems,
Journal of Global Optimization, vol. 10, pp. 381390, 1997.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
324 References
325 References
162 M. Vu, N. Devroye, M. Sharif, and V. Tarokh, Scaling laws of cognitive networks, in
Proc. ICST International Conference on Cognitive Radio Oriented Wireless Networks
and Communications (CrownCom), pp. 28, Orlando, FL, July 31August 3, 2007.
163 W. Wang, V. Srinivasan, and K.C. Chua, Using mobile relays to prolong the lifetime of
wireless sensor networks, in Proc. ACM MobiCom, Cologne, pp. 270283, Germany,
August 2005.
164 W. Wang, V. Srinivasan, and K.C. Chua, Extending the lifetime of wireless sensor net-
works through mobile relays, IEEE/ACM Transactions on Networking, vol. 16, no. 5,
pp. 11081120, October 2008.
165 Z. Wang, H.R. Sajadpour, and J.J. Garcia-Luna-Aceves, The capacity and energy effi-
ciency of wireless ad hoc networks with multi-packet reception, in Proc. ACM Mobi-
Hoc, pp. 179188, Hong Kong, China, May 2630, 2008.
166 R. Wattenhofer, L. Li, P. Bahl, and Y.-M. Wang, Distributed topology control for power
efficient operation in multihop wireless ad hoc networks, in Proc. IEEE INFOCOM,
pp. 13881397, Anchorage, AK, April 2226, 2001.
167 E. Welzl, Smallest enclosing disks, Lecture Notes in Computer Science (LNCS),
vol. 555, pp. 359370, 1991.
168 D.B. West, Introduction to Graph Theory, Upper Saddle River, NJ, Prentice Hall, 2001.
169 J. Winters, On the capacity of radio communication systems with diversity in a
Rayleigh fading environment, IEEE Journal on Selected Areas in Communications,
vol. 5, no. 5, pp. 871878, June 1987.
170 J. H. Winters, Smart antenna techniques and their application to wireless ad hoc
networks, IEEE Wireless Communications Magazine, vol. 13, no. 4, pp. 7783,
August 2006.
171 WiTricity Corp., http://www.witricity.com.
172 Wireless Power Consortium, http://www.wirelesspowerconsortium.com/.
173 A. Wyglinski, M. Nekovee, and Y.T. Hou (eds.), Cognitive Radio Communications and
Networks: Principles and Practice, Elsevier, December 2009.
174 C. Xin, B. Xie, and C.-C. Shen, A novel layered graph model for topology formation
and routing in dynamic spectrum access networks, in Proc. IEEE DySPAN, pp. 308
317, Baltimore, MD, November 811, 2005.
175 S. Ye and R. S. Blum, Optimized signaling for MIMO interference systems with
feedback, IEEE Transactions on Signal Processing, vol. 51, no. 11, pp. 28392848,
November 2003.
176 C. Yin, L. Gao, and S. Cui, Scaling laws of overlaid wireless networks: a cognitive
radio network vs. a primary network, IEEE/ACM Transactions on Networking, vol. 18,
no. 4, pp. 13171329, August 2010.
177 H. Zhang and J.C. Hou, Capacity of wireless ad-hoc networks under ultra wide band
with power constraint, in Proc. IEEE INFOCOM, pp. 455465, Miami, FL, March 13
17, 2005.
178 F. Zhang, X. Liu, S.A. Hackworth, R.J. Sclabassi, and M. Sun, In vitro and in
vivo studies on wireless powering of medical sensors and implantable devices, in
Proc. IEEE/NIH Life Science Systems and Applications Workshop (LiSSA), pp. 8487,
Bethesda, MD, April 2009.
179 J. Zhao, H. Zheng, and G. Yang, Distributed coordination in dynamic spectrum alloca-
tion networks, in Proc. IEEE DySPAN, pp. 259268, Baltimore, MD, November 811,
2005.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
326 References
180 Y. Zhao, R.S. Adve, and T.J. Lim, Improving amplify-and-forward relay networks:
optimal power allocation versus selection, in Proc. IEEE International Symposium on
Information Theory, pp. 12341238, Seattle, WA, July 914, 2006.
181 R. Zheng, Information dissemination in power-constrained wireless network, in
Proc. IEEE INFOCOM, Barcelona, Catalunya, Spain, April 2329, 2006.
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Index
327
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
328 Index
core optimization space, 5, 125, 134, 146 generalized physical model, 7, 285
core optimization variable, 125 physical model, 7, 285
cross-layer convex optimization, 3 protocol model, 7, 285
cross-layer optimization, 3, 45, 119, 129, 263 signal-to-interference-and-noise ratio
crossover, 262, 271 (SINR) model, 5, 125
cutting-plane method, 4, 41, 49 interference range, 103, 247, 254, 285, 287
interior, 307
decision variable, 11, 247, 251 interior point method, 12, 39
decode-and-forward (DF), 64 iterative improvement, 260
decomposition, 4, 41
deflected (sub)gradient method, 39 Lagrangian dual problem, 39, 47, 55
determinant function, 64 Lagrangian dual variable, 47
dichotomous branching, 96 Lagrangian function, 40
double description (DD) video, 263 Lagrangian multiplier, 39, 47
dual decomposition, 46 lexicographic max-min (LMM), 3, 14, 19
dual grid, 306 linear approximation, 5, 148, 151, 179
dual problem, 4, 12, 47 grid linearization, 148
dual simplex method, 12 piecewise linear approximation, 150, 152,
dual variable, 39 167
duality relationship, 12 linear marking, 72, 75
linear programming (LP), 3, 11, 13
empty cell, 290, 304 dual problem, 12
energy-harvesting, 151 dual simplex method, 12
entry point, 309
parametric analysis (PA), 14, 22
expectation function, 64
standard form, 12
linear relaxation, 5, 48, 111, 122, 152
fathoming a subproblem, 99
convex hull relaxation, 111
fictitious cost point (FCP), 6, 193, 201
LMM node lifetime, 14, 25
fitness function, 261, 270
mirror relationship, 14, 27
flow routing, 3, 14, 101, 125, 152, 194, 212,
LMM rate allocation, 14, 19
244
mirror relationship, 14, 27
generalized physical model, 7, 284, 285, 300 LMM-optimal solution, 19
genetic algorithm (GA), 7, 261 local optimal solution, 39
crossover, 262, 271 local search, 4, 96, 113, 135
fitness function, 261 basic local search, 7, 260
mutation, 262, 272 location-dependent problem, 6, 212, 215, 223,
Tournament selection, 271 241
global optimal solution, 39, 99 LP duality theory, 51
gradient method, 39
conjugate (sub)gradient method, 39 magnetic resonant coupling, 151
deflected (sub)gradient method, 39 max-min rate allocation, 14, 19
gradient projection method, 39 metaheuristic, 6, 260, 263, 277
gradient projection method, 39 genetic algorithm (GA), 7, 261
grid linearization, 148 iterative improvement, 260
simulated annealing (SA), 7 , 260, 263
Hamiltonian cycle, 152, 162, 174 tabu search (TS), 7, 261, 263
hidden terminal, 248 MIMO link, 43
highway path, 301, 302 MIMO network, 4, 40, 44, 57
highway system, 301, 302, 304, 309 mirror relationship, 14, 27
mixed-integer linear programming (MILP), 4,
identity matrix, 42, 64 13, 39, 68, 243
increasing event, 307 mixed-integer nonlinear programming
inner product, 42 (MINLP), 5, 110, 125, 251
integer linear program (ILP), 13 mixed-integer optimization problem, 243, 256
interference model, 7, 283, 285, 301, 313 mobile base station problem, 5, 212
9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
329 Index
330 Index